Sparse-view 3D Gaussian splatting seeks to render high-quality novel views of 3D scenes from a limited set of input images. While recent pose-free feed-forward methods leveraging pre-trained 3D priors have achieved impressive results, most of them rely on full fine-tuning of large Vision Transformer (ViT) backbones and incur substantial GPU costs. In this work, we introduce MuSASplat, a novel framework that dramatically reduces the computational burden of training pose-free feed-forward 3D Gaussian splats models with little compromise of rendering quality. Central to our approach is a lightweight Multi-Scale Adapter that enables efficient fine-tuning of ViT-based architectures with only a small fraction of training parameters. This design avoids the prohibitive GPU overhead associated with previous full-model adaptation techniques while maintaining high fidelity in novel view synthesis, even with very sparse input views. In addition, we introduce a Feature Fusion Aggregator that integrates features across input views effectively and efficiently. Unlike widely adopted memory banks, the Feature Fusion Aggregator ensures consistent geometric integration across input views and meanwhile mitigates the memory usage, training complexity, and computational costs significantly. Extensive experiments across diverse datasets show that MuSASplat achieves state-of-the-art rendering quality but has significantly reduced parameters and training resource requirements as compared with existing methods.

Severe weather restoration models often face the simultaneous interaction of multiple degradations in real-world scenarios. Existing approaches typically handle single or composite degradations based on scene descriptors derived from text or image embeddings. However, due to the varying proportions of different degradations within an image, these scene descriptors may not accurately differentiate between degradations, leading to suboptimal restoration in practical applications. To address this issue, we propose a novel Transformer-based restoration framework, AllRestorer, for dealing with four physical severe weather impairments: low-light, haze, rain, and snow. In AllRestorer, we enable the model to adaptively consider all weather impairments, thereby avoiding errors from scene descriptor misdirection. Specifically, we introduce the All-in-One Transformer Block (AiOTB), the core innovation of which is the ability to adaptively handle multiple degradations in a single image, beyond the limitation of existing Transformers that can only handle one type of degradation at a time. To accurately address different variations potentially present within the same type of degradation and minimize ambiguity, AiOTB utilizes a Composite Scene Embedding consisting of both image and text embeddings to define the degradation. Moreover, AiOTB includes an adaptive weight for each degradation, allowing for precise control of the restoration intensity. By leveraging AiOTB, AllRestorer avoids misdirection caused by inaccurate scene descriptors, achieving a 5.00 dB increase in PSNR compared to the baseline on the CDD-11 dataset.

Data is the foundation for the development of computer vision, and the establishment of datasets plays an important role in advancing the techniques of fine-grained visual categorization (FGVC). In the existing FGVC datasets used in computer vision, it is generally assumed that each collected instance has fixed characteristics and the distribution of different categories is relatively balanced. In contrast, the real world scenario reveals the fact that the characteristics of instances tend to vary with time and exhibit a long-tailed distribution. Hence, the collected datasets may mislead the optimization of the fine-grained classifiers, resulting in unpleasant performance in real applications. Starting from the real-world conditions and to promote the practical progress of fine-grained visual categorization, we present a Concept Drift and Long-Tailed Distribution (CDLT) dataset. Specifically, the dataset is collected by gathering 11195 images of 250 instances in different species for 47 consecutive months in their natural contexts. The collection process involves dozens of crowd workers for photographing and domain experts for labeling. Meanwhile, we propose a feature recombination framework to address the learning challenges associated with CDLT. Experimental results validate the efficacy of our method while also highlighting the limitations of popular large vision-language models (e.g., CLIP) in the context of long-tailed distributions. This emphasizes the significance of CDLT as a benchmark for investigating these challenges.

In the context of zero-shot learning, retrieving natural images using sketch queries is referred to as Zero-Shot Sketch-Based Image Retrieval (ZS-SBIR). The core problems of ZS-SBIR arise from the significant domain gap between sketches and photographs. The knowledge gap between seen and unseen categories is a crucial problems as well. To address these two issues, we propose a modality-aware bidirectional mining strategy aimed at enhancing the model’s understanding of multimodal data and mitigating the modality gap between sketches and photographs. Additionally, we introduce an improved Mixed Sample Data Augmentation (MSDA) technique to generate mixed samples as supplementary training data. Based on these generated mixed samples, we propose a mixed embedding mining strategy. This strategy aims to optimize the embedding space and enhance the model’s generalization ability from seen to unseen categories. Extensive experimental results on the TU-Berlin Ext., Sketchy Ext., and QuickDraw Ext. datasets demonstrate that our proposed model outperforms existing methods.

The goal of compositional zero-shot learning (CZSL) is to train a model to recognize images containing known attribute-object pairs. This reduces the reliance on extensive training data and enables the model to identify unseen combinations. Current CZSL methods face several challenges, including multiple attributes for a single object, disconnected training and test sets, long-tailed distribution of visual categories, and substantial differences in state representation between different objects. These factors collectively impede the precise identification of new combinations. In response to these challenges, we propose a Semi-Negative Contrastive Subclass Discriminative Network (SN-CSDN) based on contrastive learning. Firstly, we propose a semi-negative sampling strategy that incorporates carefully selected negative samples into the training process. This approach enables the model to effectively distinguish between different classes while enhancing its ability to capture fine-grained subclass features. By improving the model's sensitivity to inter-class differences and refining its recognition of subtle intra-class variations, this strategy significantly boosts overall discrimination performance. Additionally, we introduce a decoupled network branch designed to capture the intricate relationships between attributes and objects by generating more representative compositional embeddings. This branch leverages subclass information to ensure an accurate classification of synthesized embeddings while preserving the inherent visual distinctions of the original decoupled embeddings across different combinations. By improving feature representation capacity and mitigating sample imbalance, this design effectively improves model performance in long-tailed distributions. Our method has been comprehensively evaluated on three benchmark datasets, with results showing significant performance improvements that demonstrate the method's effectiveness and reliability.

Multi-focus image fusion (MFIF) addresses the challenge of partial focus by integrating multiple source images taken at different focal depths. Unlike most existing methods that rely on complex loss functions or large-scale synthetic datasets, this study approaches MFIF from a novel perspective: optimizing the input space. The core idea is to construct a high-quality MFIF input space in a cost-effective manner by using intermediate features from well-trained, non-MFIF networks. To this end, we propose a cascaded framework comprising two feature extractors, a Feature Distillation and Fusion Module (FDFM), and a focus segmentation network Y ${}^{U}$ Net. Based on our observation that discrepancy and edge features are essential for MFIF, we select a image deblurring network and a salient object detection network as feature extractors. To transform these extracted features into an MFIF-suitable input space, we propose FDFM as a training-free feature adapter. To make FDFM compatible with high-dimensional feature maps, we extend the manifold theory from the edge-preserving field and design a novel isometric domain transformation. Extensive experiments on six benchmark datasets show that 1) our model consistently outperforms 13 state-of-the-art methods in both qualitative and quantitative evaluations, and 2) the constructed input space can directly enhance the performance of many MFIF models without additional requirements.

Zero-shot Learning (ZSL) aims to enable image classifiers to recognize images from unseen classes that were not included during training. Unlike traditional supervised classification, ZSL typically relies on learning a mapping from visual features to predefined, human-understandable class concepts. While ZSL models promise to improve generalization and interpretability, their robustness under systematic input perturbations remain unclear. In this study, we present an empirical analysis about the robustness of existing ZSL methods at both class-level and concept-level. Specifically, we successfully disrupted their class prediction by the well-known non-target class attack (clsA). However, in the Generalized Zero-shot Learning (GZSL) setting, we observe that the success of clsA is only at the original best-calibrated point. After the attack, the optimal best-calibration point shifts, and ZSL models maintain relatively strong performance at other calibration points, indicating that clsA results in a spurious attack success in the GZSL. To address this, we propose the Class-Bias Enhanced Attack (CBEA), which completely eliminates GZSL accuracy across all calibrated points by enhancing the gap between seen and unseen class probabilities. Next, at concept-level attack, we introduce two novel attack modes: Class-Preserving Concept Attack (CPconA) and Non-Class-Preserving Concept Attack (NCPconA). Our extensive experiments evaluate three typical ZSL models across various architectures from the past three years and reveal that ZSL models are vulnerable not only to the traditional class attack but also to concept-based attacks. These attacks allow malicious actors to easily manipulate class predictions by erasing or introducing concepts. Our findings highlight a significant performance gap between existing approaches, emphasizing the need for improved adversarial robustness in current ZSL models. Our codes are available at https://github.com/FouriYe/AttackZSL_TIP26.

Cross-domain few-shot medical image segmentation (CD-FSMIS) offers a promising and data-efficient solution for medical applications where annotations are severely scarce and multimodal analysis is required. However, existing methods typically filter out domain-specific information to improve generalization, which inadvertently limits cross-domain performance and degrades source-domain accuracy. To address this, we present Contrastive Graph Modeling (C-Graph), a framework that leverages the structural consistency of medical images as a reliable domain-transferable prior. We represent image features as graphs, with pixels as nodes and semantic affinities as edges. A Structural Prior Graph (SPG) layer is proposed to capture and transfer target-category node dependencies and enable global structure modeling through explicit node interactions. Building upon SPG layers, we introduce a Subgraph Matching Decoding (SMD) mechanism that exploits semantic relations among nodes to guide prediction. Furthermore, we design a Confusion-minimizing Node Contrast (CNC) loss to mitigate node ambiguity and subgraph heterogeneity by contrastively enhancing node discriminability in the graph space. Our method significantly outperforms prior CD-FSMIS approaches across multiple cross-domain benchmarks, achieving state-of-the-art performance while simultaneously preserving strong segmentation accuracy on the source domain. Our code is available at https://github.com/primebo1/C-Graph.

A high-performance image compression algorithm is crucial for real-time information transmission across numerous fields. Despite rapid progress in image compression, computational inefficiency and poor redundancy modeling still pose significant bottlenecks, limiting practical applications. Inspired by the effectiveness of state space models (SSMs) in capturing long-range dependencies, we leverage SSMs to address computational inefficiency in existing methods and improve image compression from multiple perspectives. In this paper, we integrate the advantages of SSMs for better efficiency-performance trade-off and propose an enhanced image compression approach through refined context modeling, which we term MambaIC. Specifically, we explore context modeling to adaptively refine the representation of hidden states. Additionally, we introduce window-based local attention into channel-spatial entropy modeling to reduce potential spatial redundancy during compression, thereby increasing efficiency. Comprehensive qualitative and quantitative results validate the effectiveness and efficiency of our approach, particularly for high-resolution image compression. Code is released at https://github.com/AuroraZengfh/MambaIC.

Anchor-based 3D Gaussian splatting (3D-GS) exploits anchor features in 3D Gaussian prediction, which has achieved impressive 3D rendering quality with reduced Gaussian redundancy. On the other hand, it often encounters the dilemma among anchor features, model size, and rendering quality – large anchor features lead to large 3D models and high-quality rendering whereas reducing anchor features degrades Gaussian attribute prediction which leads to clear artifacts in the rendered textures and geometries. We design SOGS, an anchor-based 3D-GS technique that introduces second-order anchors to achieve superior rendering quality and reduced anchor features and model size simultaneously. Specifically, SOGS incorporates covariance-based second-order statistics and correlation across feature dimensions to augment features within each anchor, compensating for the reduced feature size and improving rendering quality effectively. In addition, it introduces a selective gradient loss to enhance the optimization of scene textures and scene geometries, leading to high-quality rendering with small anchor features. Extensive experiments over multiple widely adopted benchmarks show that SOGS achieves superior rendering quality in novel view synthesis with clearly reduced model size.

Spatial intelligence is emerging as a transformative frontier in AI, yet it remains constrained by the scarcity of largescale 3D datasets. Unlike the abundant 2D imagery, acquiring 3D data typically requires specialized sensors and laborious annotation. In this work, we present a scalable pipeline that converts single-view images into comprehensive, scale- and appearance-realistic 3D representations - including point clouds, camera poses, depth maps, and pseudo-RGBD - via integrated depth estimation, camera calibration, and scale calibration. Our method bridges the gap between the vast repository of imagery and the increasing demand for spatial scene understanding. By automatically generating authentic, scale-aware 3D data from images, we significantly reduce data collection costs and open new avenues for advancing spatial intelligence. We release two generated spatial datasets, i.e., COCO-3D and Objects365-v2-3D, and demonstrate through extensive experiments that our generated data can benefit various 3D tasks, ranging from fundamental perception to MLLMbased reasoning. These results validate our pipeline as an effective solution for developing AI systems capable of perceiving, understanding, and interacting with physical environments.

COLMAP-free 3D Gaussian Splatting (3D-GS) has recently attracted increasing attention due to its remarkable performance in reconstructing high-quality 3D scenes from unposed images or videos. However, it often struggles to handle scenes with complex camera trajectories as featured by drastic rotation and translation across adjacent camera views, leading to degraded estimation of camera poses and further local minima in joint optimization of camera poses and 3D-GS. We propose PCR-GS, an innovative COLMAP-free 3DGS technique that achieves superior 3D scene modeling and camera pose estimation via camera pose co-regularization. PCR-GS achieves regularization from two perspectives. The first is feature reprojection regularization which extracts view-robust DINO features from adjacent camera views and aligns their semantic information for camera pose regularization. The second is wavelet-based frequency regularization which exploits discrepancy in high-frequency details to further optimize the rotation matrix in camera poses. Extensive experiments over multiple real-world scenes show that the proposed PCR-GS achieves superior pose-free 3D-GS scene modeling under dramatic changes of camera trajectories.

Task vectors capture how a model changes during fine-tuning by recording the difference between pre-trained and task-specific weights. The composition of task vectors, a key operator in task arithmetic, enables models to integrate knowledge from multiple tasks without incurring significant additional inference costs. In this paper, we propose variational task vector composition (VTVC), where composition coefficients are taken as latent variables and estimated in a Bayesian inference framework. Unlike previous methods that operate at the task level, our framework focuses on sample-specific composition. Motivated by the observation of structural redundancy in task vectors, we introduce a Spike-and-Slab prior that promotes sparsity and aims to preserve the most informative components. To further address the high variance and sampling inefficiency in sparse, high-dimensional spaces, we develop a gated sampling mechanism that constructs a controllable posterior by filtering the composition coefficients based on both uncertainty and importance. This yields a more stable and interpretable variational framework by deterministically selecting reliable task components, reducing sampling variance while improving transparency and generalization. Experimental results demonstrate that our method achieves state-of-the-art average performance across a diverse range of benchmarks, including image classification and natural language understanding. These findings highlight the practical value of our approach, offering a new, efficient, and effective framework for task vector composition.

Video event localization tasks include temporal action localization (TAL), sound event detection (SED) and audio-visual event localization (AVEL). Existing methods tend to over-specialize on individual tasks, neglecting the equal importance of these different events for a complete understanding of video content. In this work, we aim to develop a unified framework to solve TAL, SED and AVEL tasks together to facilitate holistic video understanding. However, it is challenging since different tasks emphasize distinct event characteristics and there are substantial disparities in existing task-specific datasets (size/domain/duration). It leads to unsatisfactory results when applying a naive multi-task strategy. To tackle the problem, we introduce UniAV, a Unified Audio-Visual perception network to effectively learn and share mutually beneficial knowledge across tasks and modalities. Concretely, we propose a unified audio-visual encoder to derive generic representations from multiple temporal scales for videos from all tasks. Meanwhile, task-specific experts are designed to capture the unique knowledge specific to each task. Besides, instead of using separate prediction heads, we develop a novel unified language-aware classifier by utilizing semantic-aligned task prompts, enabling our model to flexibly localize various instances across tasks with an impressive open-set ability to localize novel categories. Extensive experiments demonstrate that UniAV, with its unified architecture, significantly outperforms both single-task models and the naive multi-task baseline across all three tasks. It achieves superior or on-par performances compared to the state-of-the-art task-specific methods on ActivityNet 1.3, DESED and UnAV-100 benchmarks.

Compositional Zero-Shot Learning (CZSL) aims to recognize novel compositions of seen primitives. Prior studies have attempted to either learn primitives individually (non-connected) or establish dependencies among them in the composition (fully-connected). In contrast, human comprehension of composition diverges from the aforementioned methods as humans possess the ability to make composition-aware adaptation for these primitives, instead of inferring them rigidly through the aforementioned methods. However, developing a comprehension of compositions akin to human cognition proves challenging within the confines of real space. This arises from the limitation of real-space-based methods, which often categorize attributes, objects, and compositions using three independent measures, without establishing a direct dynamic connection. To tackle this challenge, we expand the CZSL distance metric scheme to encompass complex spaces to unify the independent measures, and we establish an imaginary-connected embedding in complex space to model human understanding of attributes. To achieve this representation, we introduce an innovative visual bias-based attribute extraction module that selectively extracts attributes based on object prototypes. As a result, we are able to incorporate phase information in training and inference, serving as a metric for attribute-object dependencies while preserving the independent acquisition of primitives. We evaluate the effectiveness of our proposed approach on three benchmark datasets, illustrating its superiority compared to baseline methods.

In this work, we propose a method that leverages CLIP feature distillation, achieving efficient 3D segmentation through language guidance. Unlike previous methods that rely on multi-scale CLIP features and are limited by processing speed and storage requirements, our approach aims to streamline the workflow by directly and effectively distilling dense CLIP features, thereby achieving precise segmentation of 3D scenes using text. To achieve this, we introduce an adapter module and mitigate the noise issue in the dense CLIP feature distillation process through a self-cross-training strategy. Moreover, to enhance the accuracy of segmentation edges, this work presents a low-rank transient query attention mechanism. To ensure the consistency of segmentation for similar colors under different viewpoints, we convert the segmentation task into a classification task through label volume, which significantly improves the consistency of segmentation in color-similar areas. We also propose a simplified text augmentation strategy to alleviate the issue of ambiguity in the correspondence between CLIP features and text. Extensive experimental results show that our method surpasses current state-of-the-art technologies in both training speed and performance.

In unsupervised meta-learning, the clustering-based pseudo-labeling approach is an attractive framework, since it is model-agnostic, allowing it to synergize with supervised algorithms to learn from unlabeled data. However, the pseudo-labels suffer from clustering noise and semantic chaos problems, further impacting the effectiveness of meta-learning. In this paper, we analyze and optimize the pseudo-labeling process, including encoding and clustering, aiming to generate semantic-like pseudo-labels to narrow the gap between unsupervised and supervised meta-learning. First, during the encoding, we observe that the embedding space of existing methods lacks clustering-friendly properties, which is the primary reason for clustering noise. To address this issue, we minimize the inter-to-intra-class similarity ratio to generate clustering-friendly embedding features and validate our approach through comprehensive experiments. Then, during the clustering, we find that the semantic quality of pseudo-labels is not adequately controlled, resulting in semantic chaos of pseudo-labels. We propose a semantic-stability index to measure the semantic quality of pseudo-labels quantitatively. Based on this index, we propose the Semantic-aware Pseudo-label Reassignment mechanism to generate semantic-like pseudo-labels for all samples. Our approach is model-agnostic and can easily be integrated into existing supervised methods. To demonstrate its generalization ability, we integrate it into two representative algorithms: MAML and EP. The results on three main few-shot benchmarks clearly show that the proposed method achieves significant improvement compared to state-of-the-art models. Notably, our approach also outperforms the corresponding supervised method in three tasks.

In this paper, we propose a novel cross-attention-based generative adversarial network (GAN) for the challenging person image generation task. Cross-attention is a novel and intuitive multi-modal fusion method in which an attention/correlation matrix is calculated between two feature maps of different modalities. Specifically, we propose the novel XingGAN (or CrossingGAN), which consists of two generation branches that capture the person's appearance and shape, respectively. Moreover, we propose two novel cross-attention blocks to effectively transfer and update the person's shape and appearance embeddings for mutual improvement. This has not been considered by any other existing GAN-based image generation work. To further learn the long-range correlations between different person poses at different scales and sub-regions, we propose two novel multi-scale cross-attention blocks. To tackle the issue of independent correlation computations within the cross-attention mechanism leading to noisy and ambiguous attention weights, which hinder performance improvements, we propose a module called enhanced attention (EA). Lastly, we introduce a novel densely connected co-attention module to fuse appearance and shape features at different stages effectively. Extensive experiments on two public datasets demonstrate that the proposed method outperforms current GAN-based methods and performs on par with diffusion-based methods. However, our method is significantly faster than diffusion-based methods in both training and inference.

We propose a novel spatial-temporal graph Mamba (STG-Mamba) for the music-guided dance video synthesis task, i.e., to translate the input music to a dance video. STG-Mamba consists of two translation mappings: music-to-skeleton translation and skeleton-to-video translation. In the music-to-skeleton translation, we introduce a novel spatial-temporal graph Mamba (STGM) block to effectively construct skeleton sequences from the input music, capturing dependencies between joints in both the spatial and temporal dimensions. For the skeleton-to-video translation, we propose a novel self-supervised regularization network to translate the generated skeletons, along with a conditional image, into a dance video. Lastly, we collect a new skeleton-to-video translation dataset from the Internet, containing 54,944 video clips. Extensive experiments demonstrate that STG-Mamba achieves significantly better results than existing methods.

Sketch-based image retrieval (SBIR) relies on free-hand sketches to retrieve natural photos within the same class. However, its practical application is limited by its inability to retrieve classes absent from the training set. To address this limitation, the task has evolved into Zero-Shot Sketch-Based Image Retrieval (ZS-SBIR), where model performance is evaluated on unseen categories. Traditional SBIR primarily focuses on narrowing the domain gap between photo and sketch modalities. However, in the zero-shot setting, the model not only needs to address this cross-modal discrepancy but also requires a strong generalization capability to transfer knowledge to unseen categories. To this end, we propose a novel framework for ZS-SBIR that employs a pair-based relation-aware quadruplet loss to bridge feature gaps. By incorporating two negative samples from different modalities, the approach prevents positive features from becoming disproportionately distant from one modality while remaining close to another, thus enhancing inter-class separability. We also propose a Relation-Aware Meta-Learning Network (RAMLN) to obtain the margin, a hyper-parameter of cross-modal quadruplet loss, to improve the generalization ability of the model. RAMLN leverages external memory to store feature information, which it utilizes to assign optimal margin values. Experimental results obtained on the extended Sketchy and TU-Berlin datasets show a sharp improvement over existing state-of-the-art methods in ZS-SBIR.

With the rapid development of remote sensing technology, the range of its applications, such as urban planning and environmental monitoring, has broadened considerably. However, the vast increase in remote sensing data poses challenges for training deep neural networks, primarily due to the extensive labeled datasets required. To address this issue, zero-shot learning (ZSL) emerges as a viable approach, enabling recognition of unseen categories without relying on training samples. In this paper, we introduce a Robust and Discriminative Visual-Semantic Alignment (RDVSA) model, which utilizes attribute-guided contrastive learning for enhanced performance. By incorporating category centroids into the embedding space of visual features and attribute descriptions, our model enhances the discriminative power of learned features and stabilizes the learning process, particularly in the presence of complex data distributions and noise in remote sensing images. Additionally, we employ a Vision Transformer (ViT) and a “mean-teacher" framework to capture both global context and local details, ensuring robustness in occluded or irrelevant attribute regions. Comprehensive experiments on large-scale remote sensing benchmarks reveal that our approach surpasses current state-of-the-art methods in both ZSL and generalized zero-shot learning (GZSL) scenarios. These results underscore the effectiveness of using attribute-based contrastive learning alongside consistency constraints.

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen attribute-object compositions by leveraging prior knowledge of known primitives. However, real-world visual features of attributes and objects are often entangled, causing distribution shifts between seen and unseen combinations. Existing methods often ignore intrinsic variations and interactions among primitives, leading to poor feature discrimination and biased predictions. To address these challenges, we propose Multi-level Contextual Prototype Modulation (MCPM), a transformer-based framework with a hierarchical structure that effectively integrates attributes and objects to generate richer visual embeddings. At the feature level, we apply contrastive learning to improve discriminability across compositional tasks. At the prototype level, a subclass-driven modulator captures fine-grained attribute-object interactions, enabling better adaptation to long-tail distributions. Additionally, we introduce a Minority Attribute Enhancement (MAE) strategy that synthesizes virtual samples by mixing attribute classes, further mitigating data imbalance. Experiments on four benchmark datasets (MIT-States, C-GQA, UT-Zappos, and VAW-CZSL) show that MCPM brings significant performance improvements, verifying its effectiveness in complex composition scenes.

Referring camouflaged object detection (Ref-COD) is a recently proposed task, aiming to segment specified camouflaged objects by leveraging visual reference, i.e., a small set of referring images with salient target objects. Ref-COD poses a considerable challenge due to the difficulty of discerning camouflaged objects from their highly similar backgrounds, as well as the significant feature differences between the camouflaged objects and the provided visual reference. To tackle the above dilemma, we propose a novel uncertainty-aware transformer for the Ref-COD task, termed UAT. UAT first utilizes a cross-attention mechanism to align and integrate visual reference to guide camouflaged feature learning, and then models dependencies between patches in a probabilistic manner to learn predictive uncertainty and excavate discriminative camouflaged features. Specifically, we first design a referring feature aggregation (RFA) module to align and incorporate referring features with camouflaged features, guiding targeted specific feature learning within the feature space of camouflaged images. Then, to enhance multi-level feature extraction, we develop a cross-attention encoder (CAE) to integrate global information and multi-scale semantics between adjacent layers to excavate critical camouflage cues. More importantly, we propose a transformer probabilistic decoder (TPD) to model the dependencies between patches as Gaussian random variables to capture uncertainty-aware camouflaged features. Extensive experiments on the golden Ref-COD benchmark demonstrate the superiority of UAT over existing state-of-the-art competitors. The proposed UAT also achieves competitive performance on several conventional COD datasets, further demonstrating its scalability. The source code is available at https://github.com/CVL-hub/UAT.

Cross-domain few-shot medical image segmentation (CDFSMIS) presents the fundamental challenge of segmenting novel anatomical or tissue structures on unfamiliar medical imaging domains with limited annotated data. In this paper, we conduct an in-depth investigation of CDFSMIS and identify two critical observations: 1) the conventional matching mechanisms from existing few-shot models are particularly vulnerable to discrepancies in local characteristics between different domains and 2) the semantic representations learned from source domains often lack robustness when generalizing to unfamiliar target domains. Motivated by these insights, we propose a novel Dynamic Semantic Matching (DSM) framework that addresses these challenges through a three-component approach. First, we design a support-query feature re-weighting (SFR) mechanism that leverages multilevel hidden features to suppress domain-specific contents. Second, we introduce a dynamic semantic information selection (DSIS) strategy that adaptively identifies and combines domain-robust channels to construct generalizable representations. Third, we develop a dual-perspective semantic center calculation method to address the inherent texture imbalance in medical images. Extensive experiments on four unfamiliar target domains (MS-CMR, PI-PMR, Chest-X-Ray and ISIC2018) demonstrate that our approach significantly outperforms state-of-the-art few-shot segmentation and cross-domain few-shot segmentation models, validating the effectiveness of DSM in simultaneously addressing domain generalization and semantic matching challenges in medical image segmentation. The source code is available at https://github.com/YazhouZhu19/DSM.

Acquiring a large volume of annotated medical data is impractical due to time, financial, and legal constraints. Consequently, few-shot medical image segmentation is increasingly emerging as a prominent research direction. Nowadays, Medical scenarios pose two major challenges: 1) intra-class variation caused by diversity among support and query sets; 2) inter-class extreme imbalance resulting from background heterogeneity. However, existing prototypical networks struggle to tackle these obstacles effectively. To this end, we propose a Dual Interspersion and Flexible Deployment (DIFD) model. Drawing inspiration from military interspersion tactics, we design the dual Interspersion module to generate representative basis prototypes from support features. These basis prototypes are then deeply interacted with query features. Furthermore, we introduce a fusion factor to fuse and refine the basis prototypes. Ultimately, we seamlessly integrate and flexibly deploy the basis prototypes to facilitate correct matching between the query features and basis prototypes, thus conducive to improving the segmentation accuracy of the model. Extensive experiments on three publicly available medical image datasets demonstrate that our model significantly outshines other SoTAs (2.78% higher dice score on average across all datasets), achieving a new level of performance. The code is available at: https://github.com/zmcheng9/DIFD.

Magnetic resonance (MR) image acquisition is an inherently prolonged process, whose acceleration has long been the subject of research. This is commonly achieved by obtaining multiple undersampled images, simultaneously, through parallel imaging. In this article, we propose the dual-octave network (DONet), which is capable of learning multiscale spatial-frequency features from both the real and imaginary components of MR data, for parallel fast MR image reconstruction. More specifically, our DONet consists of a series of dual-octave convolutions (Dual-OctConvs), which are connected in a dense manner for better reuse of features. In each Dual-OctConv, the input feature maps and convolutional kernels are first split into two components (i.e., real and imaginary) and then divided into four groups according to their spatial frequencies. Then, our Dual-OctConv conducts intragroup information updating and intergroup information exchange to aggregate the contextual information across different groups. Our framework provides three appealing benefits: 1) it encourages information interaction and fusion between the real and imaginary components at various spatial frequencies to achieve richer representational capacity; 2) the dense connections between the real and imaginary groups in each Dual-OctConv make the propagation of features more efficient by feature reuse; and 3) DONet enlarges the receptive field by learning multiple spatial-frequency features of both the real and imaginary components. Extensive experiments on two popular datasets (i.e., clinical knee and fastMRI), under different undersampling patterns and acceleration factors, demonstrate the superiority of our model in accelerated parallel MR image reconstruction.

Compositional zero-shot learning (CZSL) aims to identify unobservable compositional concepts with prior knowledge of known primitives (attributes and objects). Due to distribution differences between seen and unseen components, existing methods for CZSL often ignore intrinsic variations between primitives and suffer from domain bias problems. To address this challenge, we proposed a concept-aware graph convolutional network (GCN) that utilizes cross-attentions to extract features unique to attributes and objects from paired concept-sharing inputs. The proposed model utilizes the cosine similarity between visual features and synthetic embeddings to estimate the feasibility score for each unseen composition. This score is then employed as a weight in the graph adjacency matrix. Additionally, the proposed model incorporates the Earth mover's distance (EMD) to further limit the concept of learning interest in disentanglers. Experimental results on three challenging dataset benchmarks, including UT-Zappos 50K, C-GQA, and MIT-States, demonstrate that the proposed model outperforms prior work in both closed- and open-world CZSL (OW-CZSL).

In the vanilla zero-shot learning (ZSL) paradigm, category attributes is the key for knowledge generalizable transfer from seen to unseen classes. By contrast, the current contrastive language-image pretraining (CLIP) model relies on the category names to achieve a more general ZSL-like prediction. When vanilla ZSL meets general CLIP, however, most existing methods on both sides struggle to benefit from each other. In this brief, we resort to attribute prompt tuning (APT) for improving the knowledge transferability from the pretrained CLIP model to the downstream ZSL framework for pursuing desirable feature representations. Our approach, termed as attribute prompt alignment network (APAN), leverages APT for cross-network feature alignment (CFA). In this way, we can investigate the effects of CLIP to vanilla ZSL task in the era of large model by the two branch APAN architecture. Specifically, APT takes as an input the templates of class attribute descriptions to produce attribute prompts, which are further used to both guide the localizations of visual regions across two frozen feature extraction networks, through a visual-semantic interaction attention. This enables APAN to progressively refine and align these cross-network features, thus resulting in generalizable feature representations that can capture fine-grained attribute information. For CFA, we simply introduce prediction alignment loss that constrains the predictions from these two cross-network visual features. Experimental results on three benchmark datasets well demonstrate that APAN outperforms the state-of-the-art methods by absorbing generalizable knowledge from CLIP models.

With thousands of observation satellites orbiting the Earth, massive-scale ultrahigh-definition (UHD) images are captured daily, covering vast areas of land, often extending across millions of square kilometers. These images commonly feature a wide range of ground objects, such as vehicles and rooftops, numbering from tens to hundreds. The ability to categorize the diverse types of objects in UHD aerial photographs is essential for a variety of real-world applications, including intelligent transportation systems, disaster prediction, and precision agriculture. In this study, we introduce a novel framework for categorizing UHD aerial photographs. The core of our approach is to represent the spatial configurations of ground objects topologically and encode these layouts using a binary matrix factorization (MF) technique that robustly addresses the challenge of noisy image-level labels. Specifically, for each UHD aerial photograph, we identify visually and semantically important object patches. These patches are then connected spatially to form graphlets, small graphs that capture the layout and relations between adjacent objects. To enhance the understanding of these graphlets, we propose a binary MF approach that captures their semantic content. The method integrates four key components: 1) learning binary hash codes; 2) refining noisy labels; 3) incorporating deep image-level semantics; and 4) adaptively updating the data graph. The binary MF is solved iteratively, with each graphlet being transformed into a set of discrete hash codes. These hash codes, which represent the spatial and semantic information of the graphlets, are subsequently encoded into a feature vector using a kernel machine, enabling multilabel categorization of the aerial photographs. For validation, we compiled a large-scale dataset of UHD aerial photographs, sourced from 100 of the top-ranked cities worldwide. Experimental results demonstrate that: 1) our method excels in learning categorization models from imperfect labels and 2) the integration of the four proposed attributes enables effective encoding of the graphlets into hash codes, providing a powerful representation of the UHD aerial photographs.

Monocular 3D object detection aims for precise 3D localization and identification of objects from a single-view image. Despite its recent progress, it often struggles while handling pervasive object occlusions that tend to complicate and degrade the prediction of object dimensions, depths, and orientations. We design MonoMAE, a monocular 3D detector inspired by Masked Autoencoders that addresses the object occlusion issue by masking and reconstructing objects in the feature space. MonoMAE consists of two novel designs. The first is depth-aware masking that selectively masks certain parts of non-occluded object queries in the feature space for simulating occluded object queries for network training. It masks non-occluded object queries by balancing the masked and preserved query portions adaptively according to the depth information. The second is lightweight query completion that works with the depth-aware masking to learn to reconstruct and complete the masked object queries. With the proposed feature-space occlusion and completion, MonoMAE learns enriched 3D representations that achieve superior monocular 3D detection performance qualitatively and quantitatively for both occluded and non-occluded objects. Additionally, MonoMAE learns generalizable representations that can work well in new domains.

Large-vocabulary object detectors (LVDs) aim to detect objects of many categories, which learn super objectness features and can locate objects accurately while applied to various downstream data. However, LVDs often struggle in recognizing the located objects due to domain discrepancy in data distribution and object vocabulary. At the other end, recent vision-language foundation models such as CLIP demonstrate superior open-vocabulary recognition capability. This paper presents KGD, a Knowledge Graph Distillation technique that exploits the implicit knowledge graphs (KG) in CLIP for effectively adapting LVDs to various downstream domains. KGD consists of two consecutive stages: 1) KG extraction that employs CLIP to encode downstream domain data as nodes and their feature distances as edges, constructing KG that inherits the rich semantic relations in CLIP explicitly; and 2) KG encapsulation that transfers the extracted KG into LVDs to enable accurate cross-domain object classification. In addition, KGD can extract both visual and textual KG independently, providing complementary vision and language knowledge for object localization and object classification in detection tasks over various downstream domains. Experiments over multiple widely adopted detection benchmarks show that KGD outperforms the state-of-the-art consistently by large margins. Codes will be released.

Test-time prompt tuning, which learns prompts online with unlabelled test samples during the inference stage, has demonstrated great potential by learning effective prompts on-the-fly without requiring any task-specific annotations. However, its performance often degrades clearly along the tuning process when the prompts are continuously updated with the test data flow, and the degradation becomes more severe when the domain of test samples changes continuously. We propose HisTPT, a Historical Test-time Prompt Tuning technique that memorizes the useful knowledge of the learnt test samples and enables robust test-time prompt tuning with the memorized knowledge. HisTPT introduces three types of knowledge banks, namely, local knowledge bank, hard-sample knowledge bank, and global knowledge bank, each of which works with different mechanisms for effective knowledge memorization and test-time prompt optimization. In addition, HisTPT features an adaptive knowledge retrieval mechanism that regularizes the prediction of each test sample by adaptively retrieving the memorized knowledge. Extensive experiments show that HisTPT achieves superior prompt tuning performance consistently while handling different visual recognition tasks (e.g., image classification, semantic segmentation, and object detection) and test samples from continuously changing domains.

Few-shot learning deals with the fundamental and challenging problem of learning from a few annotated samples, while being able to generalize well on new tasks. The crux of few-shot learning is to extract prior knowledge from related tasks to enable fast adaptation to a new task with a limited amount of data. In this paper, we propose meta-learning kernels with random Fourier features for few-shot learning, we call MetaKernel. Specifically, we propose learning variational random features in a data-driven manner to obtain task-specific kernels by leveraging the shared knowledge provided by related tasks in a meta-learning setting. We treat the random feature basis as the latent variable, which is estimated by variational inference. The shared knowledge from related tasks is incorporated into a context inference of the posterior, which we achieve via a long-short term memory module. To establish more expressive kernels, we deploy conditional normalizing flows based on coupling layers to achieve a richer posterior distribution over random Fourier bases. The resultant kernels are more informative and discriminative, which further improves the few-shot learning. To evaluate our method, we conduct extensive experiments on both few-shot image classification and regression tasks. A thorough ablation study demonstrates that the effectiveness of each introduced component in our method. The benchmark results on fourteen datasets demonstrate MetaKernel consistently delivers at least comparable and often better performance than state-of-the-art alternatives.

Federated learning (FL) allows multiple clients to collaboratively learn a globally shared model through cycles of model aggregation and local model training, without the need to share data. Most existing FL methods train local models separately on different clients, and then simply average their parameters to obtain a centralized model on the server side. However, these approaches generally suffer from large aggregation errors and severe local forgetting, which are particularly bad in heterogeneous data settings. To tackle these issues, in this paper, we propose a novel FL framework that uses online Laplace approximation to approximate posteriors on both the client and server side. On the server side, a multivariate Gaussian product mechanism is employed to construct and maximize a global posterior, largely reducing the aggregation errors induced by large discrepancies between local models. On the client side, a prior loss that uses the global posterior probabilistic parameters delivered from the server is designed to guide the local training. Binding such learning constraints from other clients enables our method to mitigate local forgetting. Finally, we achieve state-of-the-art results on several benchmarks, clearly demonstrating the advantages of the proposed method.

We present a novel graph Transformer generative adversarial network (GTGAN) to learn effective graph node relations in an end-to-end fashion for challenging graph-constrained architectural layout generation tasks. The proposed graph-Transformer-based generator includes a novel graph Transformer encoder that combines graph convolutions and self-attentions in a Transformer to model both local and global interactions across connected and non-connected graph nodes. Specifically, the proposed connected node attention (CNA) and non-connected node attention (NNA) aim to capture the global relations across connected nodes and non-connected nodes in the input graph, respectively. The proposed graph modeling block (GMB) aims to exploit local vertex interactions based on a house layout topology. Moreover, we propose a new node classification-based discriminator to preserve the high-level semantic and discriminative node features for different house components. To maintain the relative spatial relationships between ground truth and predicted graphs, we also propose a novel graph-based cycle-consistency loss. Finally, we propose a novel self-guided pre-training method for graph representation learning. This approach involves simultaneous masking of nodes and edges at an elevated mask ratio (i.e., 40%) and their subsequent reconstruction using an asymmetric graph-centric autoencoder architecture. This method markedly improves the model's learning proficiency and expediency. Experiments on three challenging graph-constrained architectural layout generation tasks (i.e., house layout generation, house roof generation, and building layout generation) with three public datasets demonstrate the effectiveness of the proposed method in terms of objective quantitative scores and subjective visual realism. New state-of-the-art results are established by large margins on these three tasks.

Estimating reliable geometric model parameters from the data with severe outliers is a fundamental and important task in computer vision. This paper attempts to sample high-quality subsets and select model instances to estimate parameters in the multi-structural data. To address this, we propose an effective method called Latent Semantic Consensus (LSC). The principle of LSC is to preserve the latent semantic consensus in both data points and model hypotheses. Specifically, LSC formulates the model fitting problem into two latent semantic spaces based on data points and model hypotheses, respectively. Then, LSC explores the distributions of points in the two latent semantic spaces, to remove outliers, generate high-quality model hypotheses, and effectively estimate model instances. Finally, LSC is able to provide consistent and reliable solutions within only a few milliseconds for general multi-structural model fitting, due to its deterministic fitting nature and efficiency. Compared with several state-of-the-art model fitting methods, our LSC achieves significant superiority for the performance of both accuracy and speed on synthetic data and real images.

In the past decade, deep neural networks have achieved significant progress in point cloud learning. However, collecting large-scale precisely-annotated point clouds is extremely laborious and expensive, which hinders the scalability of existing point cloud datasets and poses a bottleneck for efficient exploration of point cloud data in various tasks and applications. Label-efficient learning offers a promising solution by enabling effective deep network training with much-reduced annotation efforts. This paper presents the first comprehensive survey of label-efficient learning of point clouds. We address three critical questions in this emerging research field: i) the importance and urgency of label-efficient learning in point cloud processing, ii) the subfields it encompasses, and iii) the progress achieved in this area. To this end, we propose a taxonomy that organizes label-efficient learning methods based on the data prerequisites provided by different types of labels. We categorize four typical label-efficient learning approaches that significantly reduce point cloud annotation efforts: data augmentation, domain transfer learning, weakly-supervised learning, and pretrained foundation models. For each approach, we outline the problem setup and provide an extensive literature review that showcases relevant progress and challenges. Finally, we share our views on the current research challenges and potential future directions.

Motion capture technology is crucial in various applications like animation, virtual reality and sports analysis. With the development of deep learning methods, significant progress has been experienced in this field, producing cost-effective and user-friendly solutions for various applications. This paper provides a comprehensive review of deep learning-based human motion capture techniques. Our review aims to bridge the gap between academic research and practical applications, providing valuable insights and guidance for researchers and practitioners in deep learning-based human motion capture. Our study puts forth a new application-oriented taxonomy that comprehensively summarises five fundamental routes of motion capture technology. In addition to that, we also delve into the research priorities linked with each route, following the structure of “hardware requirements - technical routes - datasets - evaluation metrics” and extending the necessary criteria for transferring traditional motion capture systems to deep learning-based ones. Meanwhile, for the motion capture technology, the current state of the art is reviewed, the challenges are identified, and the future directions of the research are outlined.

Most existing few-shot learning (FSL) methods require a large amount of labeled data in meta-training, which is a major limit. To reduce the requirement of labels, a semi-supervised meta-training (SSMT) setting has been proposed for FSL, which includes only a few labeled samples and numbers of unlabeled samples in base classes. However, existing methods under this setting require class-aware sample selection from the unlabeled set, which violates the assumption of unlabeled set. In this paper, we propose a practical semi-supervised meta-training setting with truly unlabeled data to facilitate the applications of FSL in realistic scenarios. To better utilize both the labeled and truly unlabeled data, we propose a simple and effective meta-training framework, called pseudo-labeling based meta-learning (PLML). Firstly, we train a classifier via common semi-supervised learning (SSL) and use it to obtain the pseudo-labels of unlabeled data. Then we build few-shot tasks from labeled and pseudo-labeled data and design a novel finetuning method with feature smoothing and noise suppression to better learn the FSL model from noise labels. Surprisingly, through extensive experiments across two FSL datasets, we find that this simple meta-training framework effectively prevents the performance degradation of various FSL models under limited labeled data, and also significantly outperforms the representative SSMT models. Besides, benefiting from meta-training, our method also improves several representative SSL algorithms as well. We provide the training code and usage examples at https://github.com/ouyangtianran/PLML.

Few-shot object detection (FSOD) identifies objects from extremely few annotated samples. Most existing FSOD methods, recently, apply the two-stage learning paradigm, which transfers the knowledge learned from abundant base classes to assist the few-shot detectors by learning the global features. However, such existing FSOD approaches seldom consider the localization of objects from local to global. Limited by the scarce training data in FSOD, the training samples of novel classes typically capture part of objects, resulting in such FSOD methods being unable to detect the completely unseen object during testing. To tackle this problem, we propose an Extensible Co-Existing Attention (ECEA) module to enable the model to infer the global object according to the local parts. Specifically, we first devise an extensible attention mechanism that starts with a local region and extends attention to co-existing regions that are similar and adjacent to the given local region. We then implement the extensible attention mechanism in different feature scales to progressively discover the full object in various receptive fields. In the training process, the model learns the extensible ability on the base stage with abundant samples and transfers it to the novel stage of continuous extensible learning, which can assist the few-shot model to quickly adapt in extending local regions to co-existing regions. Extensive experiments on the PASCAL VOC and COCO datasets show that our ECEA module can assist the few-shot detector to completely predict the object despite some regions failing to appear in the training samples and achieve the new state-of-the-art compared with existing FSOD methods. Code is released at https://github.com/zhimengXin/ECEA.

Depth information opens up new opportunities for video object segmentation (VOS) to be more accurate and robust in complex scenes. However, the RGBD VOS task is largely unexplored due to the expensive collection of RGBD data and time-consuming annotation of segmentation. In this work, we first introduce a new benchmark for RGBD VOS, named DepthVOS, which contains 350 videos (over 55k frames in total) annotated with masks and bounding boxes. We futher propose a novel, strong baseline model - Fused Color-Depth Network (FusedCDNet), which can be trained solely under the supervision of bounding boxes, while being used to generate masks with a bounding box guideline only in the first frame. Thereby, the model possesses three major advantages: a weakly-supervised training strategy to overcome the high-cost annotation, a cross-modal fusion module to handle complex scenes, and weakly-supervised inference to promote ease of use. Extensive experiments demonstrate that our proposed method performs on par with top fully-supervised algorithms. We will open-source our project on https://github.com/yjybuaa/depthvos/ to facilitate the development of RGBD VOS.

Automatic medical image segmentation has witnessed significant development with the success of large models on massive datasets. However, acquiring and annotating vast medical image datasets often proves to be impractical due to the time consumption, specialized expertise requirements, and compliance with patient privacy standards, etc. As a result, Few-shot Medical Image Segmentation (FSMIS) has become an increasingly compelling research direction. Conventional FSMIS methods usually learn prototypes from support images and apply nearest-neighbor searching to segment the query images. However, only a single prototype cannot well represent the distribution of each class, thus leading to restricted performance. To address this problem, we propose to Generate Multiple Representative Descriptors (GMRD), which can comprehensively represent the commonality within the corresponding class distribution. In addition, we design a Multiple Affinity Maps based Prediction (MAMP) module to fuse the multiple affinity maps generated by the aforementioned descriptors. Furthermore, to address intra-class variation and enhance the representativeness of descriptors, we introduce two novel losses. Notably, our model is structured as a dual-path design to achieve a balance between foreground and background differences in medical images. Extensive experiments on four publicly available medical image datasets demonstrate that our method outperforms the state-of-the-art methods, and the detailed analysis also verifies the effectiveness of our designed module.

Zero-shot learning (ZSL) tackles the unseen class recognition problem by transferring semantic knowledge from seen classes to unseen ones. Typically, to guarantee desirable knowledge transfer, a direct embedding is adopted for associating the visual and semantic domains in ZSL. However, most existing ZSL methods focus on learning the embedding from implicit global features or image regions to the semantic space. Thus, they fail to: 1) exploit the appearance relationship priors between various local regions in a single image, which corresponds to the semantic information and 2) learn cooperative global and local features jointly for discriminative feature representations. In this article, we propose the novel graph navigated dual attention network (GNDAN) for ZSL to address these drawbacks. GNDAN employs a region-guided attention network (RAN) and a region-guided graph attention network (RGAT) to jointly learn a discriminative local embedding and incorporate global context for exploiting explicit global embeddings under the guidance of a graph. Specifically, RAN uses soft spatial attention to discover discriminative regions for generating local embeddings. Meanwhile, RGAT employs an attribute-based attention to obtain attribute-based region features, where each attribute focuses on the most relevant image regions. Motivated by the graph neural network (GNN), which is beneficial for structural relationship representations, RGAT further leverages a graph attention network to exploit the relationships between the attribute-based region features for explicit global embedding representations. Based on the self-calibration mechanism, the joint visual embedding learned is matched with the semantic embedding to form the final prediction. Extensive experiments on three benchmark datasets demonstrate that the proposed GNDAN achieves superior performances to the state-of-the-art methods. Our code and trained models are available at https://github.com/shiming-chen/GNDAN.

Super-resolving the magnetic resonance (MR) image of a target contrast under the guidance of the corresponding auxiliary contrast, which provides additional anatomical information, is a new and effective solution for fast MR imaging. However, current multi-contrast super-resolution (SR) methods tend to concatenate different contrasts directly, ignoring their relationships in different clues, e.g., in the high-and low-intensity regions. In this study, we propose a separable attention network (comprising high-intensity priority (HP) attention and low-intensity separation (LS) attention), named SANet. Our SANet could explore the areas of high-and low-intensity regions in the "forward" and "reverse" directions with the help of the auxiliary contrast while learning clearer anatomical structure and edge information for the SR of a target-contrast MR image. SANet provides three appealing benefits: First, it is the first model to explore a separable attention mechanism that uses the auxiliary contrast to predict the high-and low-intensity regions, diverting more attention to refining any uncertain details between these regions and correcting the fine areas in the reconstructed results. Second, a multistage integration module is proposed to learn the response of multi-contrast fusion at multiple stages, get the dependency between the fused representations, and boost their representation ability. Third, extensive experiments with various state-of-the-art multi-contrast SR methods on fastMRI and clinical in vivo datasets demonstrate the superiority of our model. The code is released at https://github.com/chunmeifeng/SANet.

Recognizing human-object interaction (HOI) aims at inferring various relationships between actions and objects. Although great progress in HOI has been made, the long-tail problem and combinatorial explosion problem are still practical challenges. To this end, we formulate HOI as a few-shot task to tackle both challenges and design a novel dynamic generation method to address this task. The proposed approach is called semantic-aware dynamic generation networks (SADG-Nets). Specifically, SADG-Net first assigns semantic-aware task representations for different batches of data, which further generates dynamic parameters. It obtains the features that highlight intercategory discriminability and intracategory commonality adaptively. In addition, we also design a dual semantic-aware encoder module (DSAE-Module), that is, verb-aware and noun-aware branches, to yield both action and object prototypes of HOI for each task space, which generalizes to novel combinations by transferring similarities among interactions. Extensive experimental results on two benchmark datasets, that is, humans interacting with common objects (HICO)-FS and trento universal HOI (TUHOI)-FS, illustrate that our SADG-Net achieves superior performance over state-of-the-art approaches, which proves its impressive effectiveness on few-shot HOI recognition.

Recently, deep metric learning (DML) has achieved great success. Some existing DML methods propose adaptive sample mining strategies, which learn to weight the samples, leading to interesting performance. However, these methods suffer from a small memory (e.g., one training batch), limiting their efficacy. In this work, we introduce a data-driven method, meta-mining strategy with semiglobal information (MMSI), to apply meta-learning to learn to weight samples during the whole training, leading to an adaptive mining strategy. To introduce richer information than one training batch only, we elaborately take advantage of the validation set of meta-learning by implicitly adding additional validation sample information to training. Furthermore, motivated by the latest self-supervised learning, we introduce a dictionary (memory) that maintains very large and diverse information. Together with the validation set, this dictionary presents much richer information to the training, leading to promising performance. In addition, we propose a new theoretical framework that can formulate pairwise and tripletwise metric learning loss functions in a unified framework. This framework brings new insights to society and facilitates us to generalize our MMSI to many existing DML methods. We conduct extensive experiments on three public datasets, CUB200-2011, Cars-196, and Stanford Online Products (SOP). Results show that our method can achieve the state of the art or very competitive performance. Our source codes have been made available at https://github.com/NUST-Machine-Intelligence-Laboratory/MMSI.

The performance of zero-shot learning (ZSL) can be improved progressively by learning better features and generating pseudosamples for unseen classes. Existing ZSL works typically learn feature extractors and generators independently, which may shift the unseen samples away from their real distribution and suffers from the domain bias problem. In this article, to tackle this challenge, we propose a variational autoencoder (VAE)-based framework, that is, joint Attentive Region Embedding with Enhanced Semantics (AREES), which is tailored to advance the zero-shot recognition. Specifically, AREES is end-to-end trainable and consists of three network branches: 1) attentive region embedding is used to learn the semantic-guided visual features by the attention mechanism (AM); 2) a decomposition structure and a semantic pivot regularization are used to extract enhanced semantics; and 3) a multimodal VAE (mVAE) with the cross-reconstruction loss and the distribution alignment loss is used to obtain a shared latent embedding space of visual features and semantics. Finally, features' extraction and features' generation are optimized together in AREES to address the domain shift problem to a large extent. The comprehensive evaluations on six benchmarks, including the ImageNet, demonstrate the superiority of the proposed model over its state-of-the-art counterparts.

Spotting camouflaged objects that are visually assimilated into the background is tricky for both object detection algorithms and humans who are usually confused or cheated by the perfectly intrinsic similarities between the foreground objects and the background surroundings. To tackle this challenge, we aim to extract the high-resolution texture details to avoid the detail degradation that causes blurred vision in edges and boundaries. We introduce a novel HitNet to refine the low-resolution representations by high-resolution features in an iterative feedback manner, essentially a global loop-based connection among the multi-scale resolutions. To design better feedback feature flow and avoid the feature corruption caused by recurrent path, an iterative feedback strategy is proposed to impose more constraints on each feedback connection. Extensive experiments on four challenging datasets demonstrate that our HitNet breaks the performance bottleneck and achieves significant improvements compared with 29 state-of-the-art methods. In addition, to address the data scarcity in camouflaged scenarios, we provide an application example to convert the salient objects to camouflaged objects, thereby generating more camouflaged training samples from the diverse salient object datasets. Code will be made publicly available.

We present a novel graph Transformer generative adversarial network (GTGAN) to learn effective graph node relations in an end-to-end fashion for the challenging graph-constrained house generation task. The proposed graph-Transformer-based generator includes a novel graph Transformer encoder that combines graph convolutions and self-attentions in a Transformer to model both local and global interactions across connected and non-connected graph nodes. Specifically, the proposed connected node attention (CNA) and non-connected node attention (NNA) aim to capture the global relations across connected nodes and non-connected nodes in the input graph, respectively. The proposed graph modeling block (GMB) aims to exploit local vertex interactions based on a house layout topology. More-over, we propose a new node classification-based discriminator to preserve the high-level semantic and discriminative node features for different house components. Finally, we propose a novel graph-based cycle-consistency loss that aims at maintaining the relative spatial relationships between ground truth and predicted graphs. Experiments on two challenging graph-constrained house generation tasks (i.e., house layout and roof generation) with two public datasets demonstrate the effectiveness of GTGAN in terms of objective quantitative scores and subjective visual realism. New state-of-the-art results are established by large margins on both tasks.

Contrastive learning has recently demonstrated great potential for unsupervised pre-training in 3D scene understanding tasks. However, most existing work randomly selects point features as anchors while building contrast, leading to a clear bias toward background points that often dominate in 3D scenes. Also, object awareness and foreground-to-background discrimination are neglected, making contrastive learning less effective. To tackle these issues, we propose a general foreground-aware feature contrast (FAC) framework to learn more effective point cloud representations in pre-training. FAC consists of two novel contrast designs to construct more effective and informative contrast pairs. The first is building positive pairs within the same foreground segment where points tend to have the same semantics. The second is that we prevent over-discrimination between 3D segments/objects and encourage foreground-to-background distinctions at the segment level with adaptive feature learning in a Siamese correspondence network, which adaptively learns feature correlations within and across point cloud views effectively. Visualization with point activation maps shows that our contrast pairs capture clear correspondences among fore-ground regions during pre-training. Quantitative experiments also show that FAC achieves superior knowledge transfer and data efficiency in various downstream 3D semantic segmentation and object detection tasks. All codes, data, and models are available.

Neural Radiance Field (NeRF) has shown impressive performance in novel view synthesis via implicit scene representation. However, it usually suffers from poor scalability as requiring densely sampled images for each new scene. Several studies have attempted to mitigate this problem by integrating Multi-View Stereo (MVS) technique into NeRF while they still entail a cumbersome fine-tuning process for new scenes. Notably, the rendering quality will drop severely without this fine-tuning process and the errors mainly appear around the high-frequency features. In the light of this observation, we design WaveNeRF, which integrates wavelet frequency decomposition into MVS and NeRF to achieve generalizable yet high-quality synthesis without any per-scene optimization. To preserve high-frequency information when generating 3D feature volumes, WaveNeRF builds Multi-View Stereo in the Wavelet domain by integrating the discrete wavelet transform into the classical cascade MVS, which disentangles high-frequency information explicitly. With that, disentangled frequency features can be injected into classic NeRF via a novel hybrid neural renderer to yield faithful high-frequency details, and an intuitive frequency-guided sampling strategy can be designed to suppress artifacts around high-frequency regions. Extensive experiments over three widely studied benchmarks show that WaveNeRF achieves superior generalizable radiance field modeling when only given three images as input.

Pose-free neural radiance fields (NeRF) aim to train NeRF with unposed multi-view images and it has achieved very impressive success in recent years. Most existing works share the pipeline of training a coarse pose estimator with rendered images at first, followed by a joint optimization of estimated poses and neural radiance field. However, as the pose estimator is trained with only rendered images, the pose estimation is usually biased or inaccurate for real images due to the domain gap between real images and rendered images, leading to poor robustness for the pose estimation of real images and further local minima in joint optimization. We design IR-NeRF, an innovative pose-free NeRF that introduces implicit pose regularization to refine pose estimator with unposed real images and improve the robustness of the pose estimation for real images. With a collection of 2D images of a specific scene, IR-NeRF constructs a scene codebook that stores scene features and captures the scene-specific pose distribution implicitly as priors. Thus, the robustness of pose estimation can be promoted with the scene priors according to the rationale that a 2D real image can be well reconstructed from the scene codebook only when its estimated pose lies within the pose distribution. Extensive experiments show that IR-NeRF achieves superior novel view synthesis and outperforms the state-of-the-art consistently across multiple synthetic and real datasets.

Vision-Language Pre-training has demonstrated its remarkable zero-shot recognition ability and potential to learn generalizable visual representations from languagesupervision. Taking a step ahead, language-supervised semantic segmentation enables spatial localization of textual inputs by learning pixel grouping solely from image-text pairs. Nevertheless, the state-of-the-art suffers from a clear semantic gap between visual and textual modalities: plenty of visual concepts appeared in images are missing in their paired captions. Such semantic misalignment circulates in pre-training, leading to inferior zero-shot performance in dense predictions due to insufficient visual concepts captured in textual representations. To close such semantic gap, we propose Concept Curation (CoCu), a pipeline that leverages CLIP to compensate for the missing semantics. For each image-text pair, we establish a concept archive that maintains potential visually-matched concepts with our proposed vision-driven expansion and text-to-vision-guided ranking. Relevant concepts can thus be identified via cluster-guided sampling and fed into pre-training, thereby bridging the gap between visual and textual semantics. Extensive experiments over a broad suite of 8 segmentation benchmarks show that CoCu achieves superb zero-shot transfer performance and greatly boosts language-supervised segmentation baseline by a large margin, suggesting the value of closing semantic gap in pre-training data.

Previous video object segmentation approaches mainly focus on simplex solutions linking appearance and motion, limiting effective feature collaboration between these two cues. In this work, we study a novel and efficient full-duplex strategy network (FSNet) to address this issue, by considering a better mutual restraint scheme linking motion and appearance allowing exploitation of cross-modal features from the fusion and decoding stage. Specifically, we introduce a relational cross-attention module (RCAM) to achieve bidirectional message propagation across embedding sub-spaces. To improve the model’s robustness and update inconsistent features from the spatiotemporal embeddings, we adopt a bidirectional purification module after the RCAM. Extensive experiments on five popular benchmarks show that our FSNet is robust to various challenging scenarios (e.g., motion blur and occlusion), and compares well to leading methods both for video object segmentation and video salient object detection. The project is publicly available at https://github.com/GewelsJI/FSNet .

Subspace clustering is a classical technique that has been widely used for human motion segmentation and other related tasks. However, existing segmentation methods often cluster data without guidance from prior knowledge, resulting in unsatisfactory segmentation results. To this end, we propose a novel Consistency and Diversity induced human Motion Segmentation (CDMS) algorithm. Specifically, our model factorizes the source and target data into distinct multi-layer feature spaces, in which transfer subspace learning is conducted on different layers to capture multi-level information. A multi-mutual consistency learning strategy is carried out to reduce the domain gap between the source and target data. In this way, the domain-specific knowledge and domain-invariant properties can be explored simultaneously. Besides, a novel constraint based on the Hilbert Schmidt Independence Criterion (HSIC) is introduced to ensure the diversity of multi-level subspace representations, which enables the complementarity of multi-level representations to be explored to boost the transfer learning performance. Moreover, to preserve the temporal correlations, an enhanced graph regularizer is imposed on the learned representation coefficients and the multi-level representations of the source data. The proposed model can be efficiently solved using the Alternating Direction Method of Multipliers (ADMM) algorithm. Extensive experimental results on public human motion datasets demonstrate the effectiveness of our method against several state-of-the-art approaches.

In this paper, we address the task of semantic-guided image generation. One challenge common to most existing image-level generation methods is the difficulty in generating small objects and detailed local textures. To address this, in this work we consider generating images using local context. As such, we design a local class-specific generative network using semantic maps as guidance, which separately constructs and learns subgenerators for different classes, enabling it to capture finer details. To learn more discriminative class-specific feature representations for the local generation, we also propose a novel classification module. To combine the advantages of both global image-level and local class-specific generation, a joint generation network is designed with an attention fusion module and a dual-discriminator structure embedded. Lastly, we propose a novel semantic-aware upsampling method, which has a larger receptive field and can take far-away pixels that are semantically related for feature upsampling, enabling it to better preserve semantic consistency for instances with the same semantic labels. Extensive experiments on two image generation tasks show the superior performance of the proposed method. State-of-the-art results are established by large margins on both tasks and on nine challenging public benchmarks. The source code and trained models are available at https://github.com/Ha0Tang/LGGAN.

In deformable registration, the Riemannian framework - Large Deformation Diffeomorphic Metric Mapping, or LDDMM for short - has inspired numerous techniques for comparing, deforming, averaging and analyzing shapes or images. Grounded in flows of vector fields, akin to the equations of motion used in fluid dynamics, LDDMM algorithms solve the flow equation in the space of plausible deformations, i.e., diffeomorphisms. In this work, we make use of deep residual neural networks to solve the non-stationary ODE (flow equation) based on an Euler's discretization scheme. The central idea is to represent time-dependent velocity fields as fully connected ReLU neural networks (building blocks) and derive optimal weights by minimizing a regularized loss function. Computing minimizing paths between deformations, thus between shapes, turns to find optimal network parameters by back-propagating over the intermediate building blocks. Geometrically, at each time step, our algorithm searches for an optimal partition of the space into multiple polytopes, and then computes optimal velocity vectors as affine transformations on each of these polytopes. As a result, different parts of the shape, even if they are close (such as two fingers of a hand), can be made to belong to different polytopes, and therefore be moved in different directions without costing too much energy. Importantly, we show how diffeomorphic transformations, or more precisely bilipshitz transformations, are predicted by our registration algorithm. We illustrate these ideas on diverse registration problems of 3D shapes under complex topology-preserving transformations. We thus provide essential foundations for more advanced shape variability analysis under a novel joint geometric-neural networks Riemannian-like framework, i.e., ResNet-LDDMM.

We propose a fast single-stage method for both image and video instance segmentation, called SipMask, that preserves the instance spatial information by performing multiple sub-region mask predictions. The main module in our method is a light-weight spatial preservation (SP) module that generates a separate set of spatial coefficients for the sub-regions within a bounding-box, enabling a better delineation of spatially adjacent instances. To better correlate mask prediction with object detection, we further propose a mask alignment weighting loss and a feature alignment scheme. In addition, we identify two issues that impede the performance of single-stage instance segmentation and introduce two modules, including a sample selection scheme and an instance refinement module, to address these two issues. Experiments are performed on both image instance segmentation dataset MS COCO and video instance segmentation dataset YouTube-VIS. On MS COCO test-dev set, our method achieves a state-of-the-art performance. In terms of real-time capabilities, it outperforms YOLACT by a gain of 3.0% (mask AP) under the similar settings, while operating at a comparable speed. On YouTube-VIS validation set, our method also achieves promising results. The source code is available at https://github.com/JialeCao001/SipMask.

Zero-shot learning (ZSL) tackles the novel class recognition problem by transferring semantic knowledge from seen classes to unseen ones. Semantic knowledge is typically represented by attribute descriptions shared between different classes, which act as strong priors for localizing object attributes that represent discriminative region features, enabling significant and sufficient visual-semantic interaction for advancing ZSL. Existing attention-based models have struggled to learn inferior region features in a single image by solely using unidirectional attention, which ignore the transferable and discriminative attribute localization of visual features for representing the key semantic knowledge for effective knowledge transfer in ZSL. In this paper, we propose a cross attribute-guided Transformer network, termed TransZero++, to refine visual features and learn accurate attribute localization for key semantic knowledge representations in ZSL. Specifically, TransZero++ employs an attribute → visual Transformer sub-net (AVT) and a visual → attribute Transformer sub-net (VAT) to learn attribute-based visual features and visual-based attribute features, respectively. By further introducing feature-level and prediction-level semantical collaborative losses, the two attribute-guided transformers teach each other to learn semantic-augmented visual embeddings for key semantic knowledge representations via semantical collaborative learning. Finally, the semantic-augmented visual embeddings learned by AVT and VAT are fused to conduct desirable visual-semantic interaction cooperated with class semantic vectors for ZSL classification. Extensive experiments show that TransZero++ achieves the new state-of-the-art results on three golden ZSL benchmarks and on the large-scale ImageNet dataset. The project website is available at: https://shiming-chen.github.io/TransZero-pp/TransZero-pp.html.

In this article, we provide an intuitive viewing to simplify the Siamese-based trackers by converting the tracking task to a classification. Under this viewing, we perform an in-depth analysis for them through visual simulations and real tracking examples, and find that the failure cases in some challenging situations can be regarded as the issue of missing decisive samples in offline training. Since the samples in the initial (first) frame contain rich sequence-specific information, we can regard them as the decisive samples to represent the whole sequence. To quickly adapt the base model to new scenes, a compact latent network is presented via fully using these decisive samples. Specifically, we present a statistics-based compact latent feature for fast adjustment by efficiently extracting the sequence-specific information. Furthermore, a new diverse sample mining strategy is designed for training to further improve the discrimination ability of the proposed compact latent network. Finally, a conditional updating strategy is proposed to efficiently update the basic models to handle scene variation during the tracking phase. To evaluate the generalization ability and effectiveness and of our method, we apply it to adjust three classical Siamese-based trackers, namely SiamRPN++, SiamFC, and SiamBAN. Extensive experimental results on six recent datasets demonstrate that all three adjusted trackers obtain the superior performance in terms of the accuracy, while having high running speed.

It is uncertain whether the power of transformer architectures can complement existing convolutional neural networks. A few recent attempts have combined convolution with transformer design through a range of structures in series, where the main contribution of this paper is to explore a parallel design approach. While previous transformed-based approaches need to segment the image into patch-wise tokens, we observe that the multi-head self-attention conducted on convolutional features is mainly sensitive to global correlations and that the performance degrades when these correlations are not exhibited. We propose two parallel modules along with multi-head self-attention to enhance the transformer. For local information, a dynamic local enhancement module leverages convolution to dynamically and explicitly enhance positive local patches and suppress the response to less informative ones. For mid-level structure, a novel unary co-occurrence excitation module utilizes convolution to actively search the local co-occurrence between patches. The parallel-designed Dynamic Unary Convolution in Transformer (DUCT) blocks are aggregated into a deep architecture, which is comprehensively evaluated across essential computer vision tasks in image-based classification, segmentation, retrieval and density estimation. Both qualitative and quantitative results show our parallel convolutional-transformer approach with dynamic and unary convolution outperforms existing series-designed structures.

In this paper, we identify and address a serious design bias of existing salient object detection (SOD) datasets, which unrealistically assume that each image should contain at least one clear and uncluttered salient object. This design bias has led to a saturation in performance for state-of-the-art SOD models when evaluated on existing datasets. However, these models are still far from satisfactory when applied to real-world scenes. Based on our analyses, we propose a new high-quality dataset and update the previous saliency benchmark. Specifically, our dataset, called Salient Objects in Clutter (SOC), includes images with both salient and non-salient objects from several common object categories. In addition to object category annotations, each salient image is accompanied by attributes that reflect common challenges in common scenes, which can help provide deeper insight into the SOD problem. Further, with a given saliency encoder, e.g., the backbone network, existing saliency models are designed to achieve mapping from the training image set to the training ground-truth set. We therefore argue that improving the dataset can yield higher performance gains than focusing only on the decoder design. With this in mind, we investigate several dataset-enhancement strategies, including label smoothing to implicitly emphasize salient boundaries, random image augmentation to adapt saliency models to various scenarios, and self-supervised learning as a regularization strategy to learn from small datasets. Our extensive results demonstrate the effectiveness of these tricks. We also provide a comprehensive benchmark for SOD, which can be found in our repository: https://github.com/DengPingFan/SODBenchmark.

Image complexity (IC) is an essential visual perception for human beings to understand an image. However, explicitly evaluating the IC is challenging, and has long been overlooked since, on the one hand, the evaluation of IC is relatively subjective due to its dependence on human perception, and on the other hand, the IC is semantic-dependent while real-world images are diverse. To facilitate the research of IC assessment in this deep learning era, we built the first, to our best knowledge, large-scale IC dataset with 9,600 well-annotated images. The images are of diverse areas such as abstract, paintings and real-world scenes, each of which is elaborately annotated by 17 human contributors. Powered by this high-quality dataset, we further provide a base model to predict the IC scores and estimate the complexity density maps in a weakly supervised way. The model is verified to be effective, and correlates well with human perception (with the Pearson correlation coefficient being 0.949). Last but not the least, we have empirically validated that the exploration of IC can provide auxiliary information and boost the performance of a wide range of computer vision tasks. The dataset and source code can be found at https://github.com/tinglyfeng/IC9600.

Multi-label zero-shot learning strives to classify images into multiple unseen categories for which no data is available during training. The test samples can additionally contain seen categories in the generalized variant. Existing approaches rely on learning either shared or label-specific attention from the seen classes. Nevertheless, computing reliable attention maps for unseen classes during inference in a multi-label setting is still a challenge. In contrast, state-of-the-art single-label generative adversarial network (GAN) based approaches learn to directly synthesize the class-specific visual features from the corresponding class attribute embeddings. However, synthesizing multi-label features from GANs is still unexplored in the context of zero-shot setting. When multiple objects occur jointly in a single image, a critical question is how to effectively fuse multi-class information. In this work, we introduce different fusion approaches at the attribute-level, feature-level and cross-level (across attribute and feature-levels) for synthesizing multi-label features from their corresponding multi-label class embeddings. To the best of our knowledge, our work is the first to tackle the problem of multi-label feature synthesis in the (generalized) zero-shot setting. Our cross-level fusion-based generative approach outperforms the state-of-the-art on three zero-shot benchmarks: NUS-WIDE, Open Images and MS COCO. Furthermore, we show the generalization capabilities of our fusion approach in the zero-shot detection task on MS COCO, achieving favorable performance against existing methods.

Normalization techniques are essential for accelerating the training and improving the generalization of deep neural networks (DNNs), and have successfully been used in various applications. This paper reviews and comments on the past, present and future of normalization methods in the context of DNN training. We provide a unified picture of the main motivation behind different approaches from the perspective of optimization, and present a taxonomy for understanding the similarities and differences between them. Specifically, we decompose the pipeline of the most representative normalizing activation methods into three components: the normalization area partitioning, normalization operation and normalization representation recovery. In doing so, we provide insight for designing new normalization technique. Finally, we discuss the current progress in understanding normalization methods, and provide a comprehensive review of the applications of normalization for particular tasks, in which it can effectively solve the key issues.

Automatic modulation classification (AMC) is an important technology for the monitoring, management, and control of communication systems. In recent years, machine learning approaches are becoming popular to improve the effectiveness of AMC for radio signals. However, the automatic modulation open-set recognition (AMOSR) scheme that aims to identify the known modulation types and recognize the unknown modulation signals is not well studied. Therefore, in this paper, we propose a novel multi-modal marginal prototype framework for radio frequency (RF) signals (MMPRF) to improve AMOSR performance. First, MMPRF addresses the problem of simultaneous recognition of closed and open sets by partitioning the feature space in the way of one versus other and marginal restrictions. Second, we exploit the wireless signal domain knowledge to extract a series of signal-related features to enhance the AMOSR capability. In addition, we propose a GAN-based unknown sample generation strategy to allow the model to understand the unknown world. Finally, we conduct extensive experiments on several publicly available radio modulation data, and experimental results show that our proposed MMPRF outperforms the state-of-the-art AMOSR methods.

Binary neural networks (BNNs) have attracted broad research interest due to their efficient storage and computational ability. Nevertheless, a significant challenge of BNNs lies in handling discrete constraints while ensuring bit entropy maximization, which typically makes their weight optimization very difficult. Existing methods relax the learning using the sign function, which simply encodes positive weights into +1s, and -1s otherwise. Alternatively, we formulate an angle alignment objective to constrain the weight binarization to$\lbrace$0,+1$\rbrace$to solve the challenge. In this paper, we show that our weight binarization provides an analytical solution by encoding high-magnitude weights into +1s, and 0 s otherwise. Therefore, a high-quality discrete solution is established in a computationally efficient manner without the sign function. We prove that the learned weights of binarized networks roughly follow a Laplacian distribution that does not allow entropy maximization, and further demonstrate that it can be effectively solved by simply removing the$\ell _{2}$regularization during network training. Our method, dubbed sign-to-magnitude network binarization (SiMaN), is evaluated on CIFAR-10 and ImageNet, demonstrating its superiority over the sign-based state-of-the-arts. Our source code, experimental settings, training logs and binary models are available athttps://github.com/lmbxmu/SiMaN.

Point cloud data have been widely explored due to its superior accuracy and robustness under various adverse situations. Meanwhile, deep neural networks (DNNs) have achieved very impressive success in various applications such as surveillance and autonomous driving. The convergence of point cloud and DNNs has led to many deep point cloud models, largely trained under the supervision of large-scale and densely-labelled point cloud data. Unsupervised point cloud representation learning, which aims to learn general and useful point cloud representations from unlabelled point cloud data, has recently attracted increasing attention due to the constraint in large-scale point cloud labelling. This paper provides a comprehensive review of unsupervised point cloud representation learning using DNNs. It first describes the motivation, general pipelines as well as terminologies of the recent studies. Relevant background including widely adopted point cloud datasets and DNN architectures is then briefly presented. This is followed by an extensive discussion of existing unsupervised point cloud representation learning methods according to their technical approaches. We also quantitatively benchmark and discuss the reviewed methods over multiple widely adopted point cloud datasets. Finally, we share our humble opinion about several challenges and problems that could be pursued in the future research in unsupervised point cloud representation learning.

Learning to re-identify or retrieve a group of people across non-overlapped camera systems has important applications in video surveillance. However, most existing methods focus on (single) person re-identification (re-id), ignoring the fact that people often walk in groups in real scenarios. In this work, we take a step further and consider employing context information for identifying groups of people, i.e., group re-id. On the one hand, group re-id is more challenging than single person re-id, since it requires both a robust modeling of local individual person appearance (with different illumination conditions, pose/viewpoint variations, and occlusions), as well as full awareness of global group structures (with group layout and group member variations). On the other hand, we believe that person re-id can be greatly enhanced by incorporating additional visual context from neighboring group members, a task which we formulate as group-aware (single) person re-id. In this paper, we propose a novel unified framework based on graph neural networks to simultaneously address the above two group-based re-id tasks, i.e., group re-id and group-aware person re-id. Specifically, we construct a context graph with group members as its nodes to exploit dependencies among different people. A multi-level attention mechanism is developed to formulate both intra-group and inter-group context, with an additional self-attention module for robust graph-level representations by attentively aggregating node-level features. The proposed model can be directly generalized to tackle group-aware person re-id using node-level representations. Meanwhile, to facilitate the deployment of deep learning models on these tasks, we build a new group re-id dataset which contains more than 3.8K images with 1.5K annotated groups, an order of magnitude larger than existing group re-id datasets. Extensive experiments on the novel dataset as well as three existing datasets clearly demonstrate the effectiveness of the proposed framework for both group-based re-id tasks.

Given a degraded input image, image restoration aims to recover the missing high-quality image content. Numerous applications demand effective image restoration, e.g., computational photography, surveillance, autonomous vehicles, and remote sensing. Significant advances in image restoration have been made in recent years, dominated by convolutional neural networks (CNNs). The widely-used CNN-based methods typically operate either on full-resolution or on progressively low-resolution representations. In the former case, spatial details are preserved but the contextual information cannot be precisely encoded. In the latter case, generated outputs are semantically reliable but spatially less accurate. This paper presents a new architecture with a holistic goal of maintaining spatially-precise high-resolution representations through the entire network, and receiving complementary contextual information from the low-resolution representations. The core of our approach is a multi-scale residual block containing the following key elements: (a) parallel multi-resolution convolution streams for extracting multi-scale features, (b) information exchange across the multi-resolution streams, (c) non-local attention mechanism for capturing contextual information, and (d) attention based multi-scale feature aggregation. Our approach learns an enriched set of features that combines contextual information from multiple scales, while simultaneously preserving the high-resolution spatial details. Extensive experiments on six real image benchmark datasets demonstrate that our method, named as MIRNet-v2, achieves state-of-the-art results for a variety of image processing tasks, including defocus deblurring, image denoising, super-resolution, and image enhancement. The source code and pre-trained models are available at https://github.com/swz30/MIRNetv2.

Although current salient object detection (SOD) works have achieved significant progress, they are limited when it comes to the integrity of the predicted salient regions. We define the concept of integrity at both a micro and macro level. Specifically, at the micro level, the model should highlight all parts that belong to a certain salient object. Meanwhile, at the macro level, the model needs to discover all salient objects in a given image. To facilitate integrity learning for SOD, we design a novel Integrity Cognition Network (ICON), which explores three important components for learning strong integrity features. 1) Unlike existing models, which focus more on feature discriminability, we introduce a diverse feature aggregation (DFA) component to aggregate features with various receptive fields (i.e., kernel shape and context) and increase feature diversity. Such diversity is the foundation for mining the integral salient objects. 2) Based on the DFA features, we introduce an integrity channel enhancement (ICE) component with the goal of enhancing feature channels that highlight the integral salient objects, while suppressing the other distracting ones. 3) After extracting the enhanced features, the part-whole verification (PWV) method is employed to determine whether the part and whole object features have strong agreement. Such part-whole agreements can further improve the micro-level integrity for each salient object. To demonstrate the effectiveness of our ICON, comprehensive experiments are conducted on seven challenging benchmarks. Our ICON outperforms the baseline methods in terms of a wide range of metrics. Notably, our ICON achieves ∼ 10% relative improvement over the previous best model in terms of average false negative ratio (FNR), on six datasets. Codes and results are available at: https://github.com/mczhuge/ICON.

Supervised machine learning methods have been widely developed for segmentation tasks in recent years. However, the quality of labels has high impact on the predictive performance of these algorithms. This issue is particularly acute in the medical image domain, where both the cost of annotation and the inter-observer variability are high. Different human experts contribute estimates of the "actual" segmentation labels in a typical label acquisition process, influenced by their personal biases and competency levels. The performance of automatic segmentation algorithms is limited when these noisy labels are used as the expert consensus label. In this work, we use two coupled CNNs to jointly learn, from purely noisy observations alone, the reliability of individual annotators and the expert consensus label distributions. The separation of the two is achieved by maximally describing the annotator's "unreliable behavior" (we call it "maximally unreliable") while achieving high fidelity with the noisy training data. We first create a toy segmentation dataset using MNIST and investigate the properties of the proposed algorithm. We then use three public medical imaging segmentation datasets to demonstrate our method's efficacy, including both simulated (where necessary) and real-world annotations: 1) ISBI2015 (multiple-sclerosis lesions); 2) BraTS (brain tumors); 3) LIDC-IDRI (lung abnormalities). Finally, we create a real-world multiple sclerosis lesion dataset (QSMSC at UCL: Queen Square Multiple Sclerosis Center at UCL, UK) with manual segmentations from 4 different annotators (3 radiologists with different level skills and 1 expert to generate the expert consensus label). In all datasets, our method consistently outperforms competing methods and relevant baselines, especially when the number of annotations is small and the amount of disagreement is large. The studies also reveal that the system is capable of capturing the complicated spatial characteristics of annotators' mistakes.

Zero-shot learning (ZSL), an emerging topic in recent years, targets at distinguishing unseen class images by taking images from seen classes for training the classifier. Existing works often build embeddings between global feature space and attribute space, which, however, neglect the treasure in image parts. Discrimination information is usually contained in the image parts, e.g., black and white striped area of a zebra is the key difference from a horse. As such, image parts can facilitate the transfer of knowledge among the seen and unseen categories. In this paper, we first conduct a brief review on ZSL with detailed descriptions of these methods. Next, to discover meaningful parts, we propose an end-to-end attention-based embedding network for ZSL, which contains two sub-streams: the attention part embedding (APE) stream, and the attention second-order embedding (ASE) stream. APE is used to discover multiple image parts based on attention. ASE is introduced for ensuring knowledge transfer stably by second-order collaboration. Furthermore, an adaptive thresholding strategy is proposed to suppress noise and redundant parts. Finally, a global branch is incorporated for the full use of global information. Experiments on four benchmarks demonstrate that our models achieve superior results under both ZSL and GZSL settings.

Zero-shot learning (ZSL) aims to classify unseen samples based on the relationship between the learned visual features and semantic features. Traditional ZSL methods typically capture the underlying multimodal data structures by learning an embedding function between the visual space and the semantic space with the Euclidean metric. However, these models suffer from the hubness problem and domain bias problem, which leads to unsatisfactory performance, especially in the generalized ZSL (GZSL) task. To tackle such a problem, we formulate a discriminative cross-aligned variational autoencoder (DCA-VAE) for ZSL. The proposed model effectively utilizes a modified cross-modal-alignment variational autoencoder (VAE) to transform both visual features and semantic features obtained by the discriminative cosine metric into latent features. The key to our method is that we collect principal discriminative information from visual and semantic features to construct latent features which contain the discriminative multimodal information associated with unseen samples. Finally, the proposed model DCA-VAE is validated on six benchmarks including the large dataset ImageNet, and several experimental results demonstrate the superiority of DCA-VAE over most existing embedding or generative ZSL models on the standard ZSL and the more realistic GZSL tasks.

Graph embedding, which aims to learn low-dimensional node representations to preserve original graph structures, has attracted extensive research interests. However, most existing graph embedding models represent nodes in Euclidean spaces, which cannot effectively preserve complex patterns, e.g., hierarchical structures. Very recently, several hyperbolic embedding models have been proposed to preserve the hierarchical information in negative curvature spaces. Nevertheless, existing hyperbolic models fail to model the asymmetric proximity between nodes. To address this, we investigate a new asymmetric hyperbolic network representation problem, which targets at jointly preserving the hierarchical structures and asymmetric proximity for general directed graphs. We solve this problem by proposing a novelRotatedLorentzianEmbedding (ROLE) model, which yields two main benefits. First, our model can effectively capture both implicit and explicit hierarchical structures that come from the network topology and category information of nodes, respectively. Second, it can model the asymmetric proximity using rotation transformations. Specifically, we represent each node with a Lorentzian embedding vector, and learn two rotation matrices to reflect the direction of edges. We conduct extensive experiments on four real-world directed graph datasets. Empirical results demonstrate that the proposed approach consistently outperforms various state-of-the-art embedding models. In particular, ROLE achieves HR@1 scores up to 19.8% higher and NDCG@5 scores up to 11.3% higher than the best baselines on the task of node recommendation.

With the proliferation of online social networks, a great deal of online user action data has been generated. Such data has enabled the study of information diffusion prediction, which is a fundamental problem for understanding the propagation of information on social media platforms. In diffusion prediction models, there are two standard components, i.e., a social graph and information diffusion cascades. We observe that both components exhibit latent hierarchical structures. However, most existing models are designed based on euclidean spaces, and hence cannot effectively capture complex patterns, especially hierarchical structures. Therefore, we investigate a novel research problem to learn hyperbolic representations for information diffusion prediction. To reflect the different characteristics of social graphs and diffusion cascades, we encode them into two latent hyperbolic spaces with different trainable curvatures. In addition, to model influence dependencies, we propose a co-attention mechanism to capture the processes of diffusion cascades using positional embeddings. Given a set of activated seed users, we jointly exploit diffusion cascades and social links to predict which users will be influenced. We conduct extensive experiments on four real-world datasets. Empirical results demonstrate that the proposed H-Diffu model significantly outperforms several state-of-the-art diffusion prediction frameworks.

Detecting the tooth-gingiva trim line from a dental surface plays a critical role in dental treatment planning and aligner 3D printing. Existing methods treat this task as a segmentation problem, which is resolved with geometric deep learning based mesh segmentation techniques. However, these methods can only provide indirect results (i.e., segmented teeth) and suffer from unsatisfactory accuracy due to the incapability of making full use of high-resolution dental surfaces. To this end, we propose a two-stage geometric deep learning framework for automatically detecting tooth-gingiva trim lines from dental surfaces. Our framework consists of a trim line proposal network (TLP-Net) for predicting an initial trim line from the low-resolution dental surface as well as a trim line refinement network (TLR-Net) for refining the initial trim line with the information from the high-resolution dental surface. Specifically, our TLP-Net predicts the initial trim line by fusing the multi-scale features from a U-Net with a proposed residual multi-scale attention fusion module. Moreover, we propose feature bridge modules and a trim line loss to further improve the accuracy. The resulting trim line is then fed to our TLR-Net, which is a deep-based LDDMM model with the high-resolution dental surface as input. In addition, dense connections are incorporated into TLR-Net for improved performance. Our framework provides an automatic solution to trim line detection by making full use of raw high-resolution dental surfaces. Extensive experiments on a clinical dental surface dataset demonstrate that our TLP-Net and TLR-Net are superior trim line detection methods and outperform cutting-edge methods in both qualitative and quantitative evaluations.

Federated learning (FL) can be used to improve data privacy and efficiency in magnetic resonance (MR) image reconstruction by enabling multiple institutions to collaborate without needing to aggregate local data. However, the domain shift caused by different MR imaging protocols can substantially degrade the performance of FL models. Recent FL techniques tend to solve this by enhancing the generalization of the global model, but they ignore the domain-specific features, which may contain important information about the device properties and be useful for local reconstruction. In this paper, we propose a specificity-preserving FL algorithm for MR image reconstruction (FedMRI). The core idea is to divide the MR reconstruction model into two parts: a globally shared encoder to obtain a generalized representation at the global level, and a client-specific decoder to preserve the domain-specific properties of each client, which is important for collaborative reconstruction when the clients have unique distribution. Such scheme is then executed in the frequency space and the image space respectively, allowing exploration of generalized representation and client-specific properties simultaneously in different spaces. Moreover, to further boost the convergence of the globally shared encoder when a domain shift is present, a weighted contrastive regularization is introduced to directly correct any deviation between the client and server during optimization. Extensive experiments demonstrate that our FedMRI's reconstructed results are the closest to the ground-truth for multi-institutional data, and that it outperforms state-of-the-art FL methods.

Accelerated multi-modal magnetic resonance (MR) imaging is a new and effective solution for fast MR imaging, providing superior performance in restoring the target modality from its undersampled counterpart with guidance from an auxiliary modality. However, existing works simply combine the auxiliary modality as prior information, lacking in-depth investigations on the potential mechanisms for fusing different modalities. Further, they usually rely on the convolutional neural networks (CNNs), which is limited by the intrinsic locality in capturing the long-distance dependency. To this end, we propose a multi-modal transformer (MTrans), which is capable of transferring multi-scale features from the target modality to the auxiliary modality, for accelerated MR imaging. To capture deep multi-modal information, our MTrans utilizes an improved multi-head attention mechanism, named cross attention module, which absorbs features from the auxiliary modality that contribute to the target modality. Our framework provides three appealing benefits: (i) Our MTrans use an improved transformers for multi-modal MR imaging, affording more global information compared with existing CNN-based methods. (ii) A new cross attention module is proposed to exploit the useful information in each modality at different scales. The small patch in the target modality aims to keep more fine details, the large patch in the auxiliary modality aims to obtain high-level context features from the larger region and supplement the target modality effectively. (iii) We evaluate MTrans with various accelerated multi-modal MR imaging tasks, e.g., MR image reconstruction and super-resolution, where MTrans outperforms state-of-the-art methods on fastMRI and real-world clinical datasets.

Zero-shot learning (ZSL) aims to recognize unknown categories that are unavailable during training. Recently, generative models have shown the potential to address this challenging problem by synthesizing unseen features conditioned on semantic embeddings such as attributes. However, unidirectional generative models cannot guarantee the effective coupling between visual and semantic spaces. To this end, we propose a visual-semantic aligned bidirectional network with cycle consistency to alleviate the gap between these two spaces, generating unseen features of high quality. More importantly, we incorporate two carefully designed strategies into our bidirectional framework to improve the overall ZSL performance. Specifically, we enhance the intra-domain class divergence in both visual and semantic spaces, and in the meantime, mitigate the inter-domain shift to preserve seen-unseen domain discrimination. Experimental results on four standard benchmarks show the superiority of our framework over existing state-of-the-art methods under both conventional and generalized ZSL settings.

Self-attention (SA) based networks have achieved great success in image captioning, constantly dominating the leaderboards of online benchmarks. However, existing SA networks still suffer from distance insensitivity and low-rank bottleneck. In this paper, we aim to optimize SA in terms of two aspects, thereby addressing the above issues. First, we introduce a Distance-sensitive Self-Attention (DSA), which considers the raw geometric distances between query-key pairs in the 2D images during SA modeling. Second, we present a simple yet effective approach, named Multi-branch Self-Attention (MSA) to compensate for the low-rank bottleneck. MSA treats a multi-head self-attention layer as a branch and duplicates it multiple times to increase the expressive power of SA. To validate the effectiveness of the two designs, we apply them to the standard self-attention network, and conduct extensive experiments on the highly competitive MS-COCO dataset. We achieve new state-of-the-art performance on both the local and online test sets,i.e., 135.1% CIDEr on the Karpathy split and 135.4% CIDEr on the official online split.

Generative Adversarial Networks (GANs) have been widely-used in image translation, but their high computational and storage costs impede the deployment on mobile devices. Prevalent methods for CNN compression cannot be directly applied to GANs due to the specificity of GAN tasks and the unstable adversarial training. To solve these, in this paper, we introduce a novel GAN compression method, termed DMAD, by proposing a Differentiable Mask and a co-Attention Distillation. The former searches for a light-weight generator architecture in a training-adaptive manner. To overcome channel inconsistency when pruning the residual connections, an adaptive cross-block group sparsity is further incorporated. The latter simultaneously distills informative attention maps from both the generator and discriminator of a pre-trained model to the searched generator, effectively stabilizing the adversarial training of our light-weight model. Experiments show that DMAD can reduce the Multiply Accumulate Operations (MACs) of CycleGAN by 13x and that of Pix2Pix by 4x while retaining a comparable performance against the full model. Our code can be available at https://github.com/SJLeo/DMAD.

Occlusion is a challenging yet commonly seen problem for facial perception. Existing works resort to deep learning models and perform model training on synthesized data due to the lack of paired real-world data. As a result,they usually perform unsatisfactorily on real-world occluded faces because of domain gaps. In this paper, we decompose the face de-occlusion task into three stages, i.e., occlusion detection, face parsing, and face reconstruction, to alleviate this issue. We first perform occlusion detection and use its results as guidance for the second stage to conduct occlusion-free face parsing. As such, face de-occlusion is first performed on the face paring space with less difficulty. We can train these two stages on both synthesized and real-world images, hence can obtain accurate results for the latter. In the last stage, we use the domain-agnostic occlusion detection map and the face parsing map as the guidance to conduct face reconstruction, thus can reduce the impact of appearance information and improve the model performance on real-world data. Aiming at improving the model capacity of inferring occluded facial appearance, we also propose two types of reference modules to use relevant facial parts to enhance the reconstruction of occluded regions. Consequently, our proposed model achieves promising face de-occlusion results on real-world images.

Unsupervised cross-domain crowd counting has recently received great attention in computer vision, which generalizes the model from the source domain to the unlabeled target domain. However, it is an extremely challenging task because only unlabeled data is available from the target domain and the domain gap between two domains is implicit in crowd counting. In this paper, we propose a latent domain generation method to improve the generalization ability of unsupervised domain adaptation crowd counting by generating a latent domain. To this end, we propose a domain generator with random perturbations to learn a new latent distribution derived from the original source distribution. The latent domain generator can extract target information sampled in its stochastic latent representation, which preserves the original target information and enhances the variational ability. Meanwhile, to ensure that the generated latent domain is consistent with the source domain in counting performance, we introduce a consistency loss to encourage similar output from latent and source domains. Moreover, to enhance the adaptation ability of the generated latent domain, we apply the adversarial loss to achieve alignment between the latent and target domains. The domain generator with the adversarial loss and consistency loss ensures that the generated domain is aligned to the target while also improving the robustness of the original source domain model. The experiment indicates that our framework can effortlessly extend to scenarios with different objects (crowd, cars). The experiments also demonstrate the effectiveness of our method on unsupervised realistic-to-realistic crowd counting problems.

Facial landmark detection is a crucial preprocessing step in many applications that process facial images. Deep-learning-based methods have become mainstream and achieved outstanding performance in facial landmark detection. However, accurate models typically have a large number of parameters, which results in high computational complexity and execution time. A simple but effective facial landmark detection model that achieves a balance between accuracy and speed is crucial. To achieve this, a lightweight, efficient, and effective model is proposed called the efficient face alignment network (EfficientFAN) in this article. EfficientFAN adopts the encoder-decoder structure, with a simple backbone EfficientNet-B0 as the encoder and three upsampling layers and convolutional layers as the decoder. Moreover, deep dark knowledge is extracted through feature-aligned distillation and patch similarity distillation on the teacher network, which contains pixel distribution information in the feature space and multiscale structural information in the affinity space of feature maps. The accuracy of EfficientFAN is further improved after it absorbs dark knowledge. Extensive experimental results on public datasets, including 300 Faces in the Wild (300W), Wider Facial Landmarks in the Wild (WFLW), and Caltech Occluded Faces in the Wild (COFW), demonstrate the superiority of EfficientFAN over state-of-the-art methods.

Existing online knowledge distillation approaches either adopt the student with the best performance or construct an ensemble model for better holistic performance. However, the former strategy ignores other students' information, while the latter increases the computational complexity during deployment. In this article, we propose a novel method for online knowledge distillation, termed feature fusion and self-distillation (FFSD), which comprises two key components: FFSD, toward solving the above problems in a unified framework. Different from previous works, where all students are treated equally, the proposed FFSD splits them into a leader student set and a common student set. Then, the feature fusion module converts the concatenation of feature maps from all common students into a fused feature map. The fused representation is used to assist the learning of the leader student. To enable the leader student to absorb more diverse information, we design an enhancement strategy to increase the diversity among students. Besides, a self-distillation module is adopted to convert the feature map of deeper layers into a shallower one. Then, the shallower layers are encouraged to mimic the transformed feature maps of the deeper layers, which helps the students to generalize better. After training, we simply adopt the leader student, which achieves superior performance, over the common students, without increasing the storage or inference cost. Extensive experiments on CIFAR-100 and ImageNet demonstrate the superiority of our FFSD over existing works. The code is available at https://github.com/SJLeo/FFSD.

This article focuses on filter-level network pruning. A novel pruning method, termed CLR-RNF, is proposed. We first reveal a "long-tail" pruning problem in magnitude-based weight pruning methods and then propose a computation-aware measurement for individual weight importance, followed by a cross-layer ranking (CLR) of weights to identify and remove the bottom-ranked weights. Consequently, the per-layer sparsity makes up the pruned network structure in our filter pruning. Then, we introduce a recommendation-based filter selection scheme where each filter recommends a group of its closest filters. To pick the preserved filters from these recommended groups, we further devise a k -reciprocal nearest filter (RNF) selection scheme where the selected filters fall into the intersection of these recommended groups. Both our pruned network structure and the filter selection are nonlearning processes, which, thus, significantly reduces the pruning complexity and differentiates our method from existing works. We conduct image classification on CIFAR-10 and ImageNet to demonstrate the superiority of our CLR-RNF over the state-of-the-arts. For example, on CIFAR-10, CLR-RNF removes 74.1% FLOPs and 95.0% parameters from VGGNet-16 with even 0.3% accuracy improvements. On ImageNet, it removes 70.2% FLOPs and 64.8% parameters from ResNet-50 with only 1.7% top-five accuracy drops. Our project is available at https://github.com/lmbxmu/CLR-RNF.

Generative (generalized) zero-shot learning [(G)ZSL] models aim to synthesize unseen class features by using only seen class feature and attribute pairs as training data. However, the generated fake unseen features tend to be dominated by the seen class features and thus classified as seen classes, which can lead to inferior performances under zero-shot learning (ZSL), and unbalanced results under generalized ZSL (GZSL). To address this challenge, we tailor a novel balanced semantic embedding generative network (BSeGN), which incorporates balanced semantic embedding learning into generative learning scenarios in the pursuit of unbiased GZSL. Specifically, we first design a feature-to-semantic embedding module (FEM) to distinguish real seen and fake unseen features collaboratively with the generator in an online manner. We introduce the bidirectional contrastive and balance losses for the FEM learning, which can guarantee a balanced prediction for the interdomain features. In turn, the updated FEM can boost the learning of the generator. Next, we propose a multilevel feature integration module (mFIM) from the cycle-consistency branch of BSeGN, which can mitigate the domain bias through feature enhancement. To the best of our knowledge, this is the first work to explore embedding and generative learning jointly within the field of ZSL. Extensive evaluations on four benchmarks demonstrate the superiority of BSeGN over its state-of-the-art counterparts.

Mesh Schelling points explain how humans focus on specific regions of a 3D object. They have a large number of important applications in computer graphics and provide valuable information for perceptual psychology studies. However, detecting mesh Schelling points is time-consuming and expensive since the existing techniques are mostly based on participant observation studies. To overcome these limitations, we propose to employ powerful deep learning techniques to detect mesh Schelling points in an automatic manner, free from participant observation studies. Specifically, we utilize the mesh convolution and pooling operations to extract informative features from mesh objects, and then predict the 3D heat map of Schelling points in an end-to-end manner. In addition, we propose a Deep Schelling Network (DS-Net) to automatically detect the Schelling points, including a multi-scale fusion component and a novel region-specific loss function to improve our network for a better regression of heat maps. To the best of our knowledge, DS-Net is the first deep neural network for detecting Schelling points from 3D meshes. We evaluate DS-Net on a mesh Schelling point dataset obtained from participant observation studies. The experimental results demonstrate that DS-Net is capable of detecting mesh Schelling points effectively and outperforms various state-of-the-art mesh saliency methods and deep learning models, both qualitatively and quantitatively.

Invariance to diverse types of image corruption, such as noise, blurring, or colour shifts, is essential to establish robust models in computer vision. Data augmentation has been the major approach in improving the robustness against common corruptions. However, the samples produced by popular augmentation strategies deviate significantly from the underlying data manifold. As a result, performance is skewed toward certain types of corruption. To address this issue, we propose a multi-source vicinal transfer augmentation (VITA) method for generating diverse on-manifold samples. The proposed VITA consists of two complementary parts: tangent transfer and integration of multi-source vicinal samples. The tangent transfer creates initial augmented samples for improving corruption robustness. The integration employs a generative model to characterize the underlying manifold built by vicinal samples, facilitating the generation of on-manifold samples. Our proposed VITA significantly outperforms the current state-of-the-art augmentation methods, demonstrated in extensive experiments on corruption benchmarks.

Semi-supervised learning is a challenging problem which aims to construct a model by learning from limited labeled examples. Numerous methods for this task focus on utilizing the predictions of unlabeled instances consistency alone to regularize networks. However, treating labeled and unlabeled data separately often leads to the discarding of mass prior knowledge learned from the labeled examples. In this paper, we propose a novel method for semi-supervised semantic segmentation named GuidedMix-Net, by leveraging labeled information to guide the learning of unlabeled instances. Specifically, GuidedMix-Net employs three operations: 1) interpolation of similar labeled-unlabeled image pairs; 2) transfer of mutual information; 3) generalization of pseudo masks. It enables segmentation models can learning the higher-quality pseudo masks of unlabeled data by transfer the knowledge from labeled samples to unlabeled data. Along with supervised learning for labeled data, the prediction of unlabeled data is jointly learned with the generated pseudo masks from the mixed data. Extensive experiments on PASCAL VOC 2012, and Cityscapes demonstrate the effectiveness of our GuidedMix-Net, which achieves competitive segmentation accuracy and significantly improves the mIoU over 7$\%$ compared to previous approaches.

Instance contrast for unsupervised representation learning has achieved great success in recent years. In this work, we explore the idea of instance contrastive learning in unsupervised domain adaptation (UDA) and propose a novel Category Contrast technique (CaCo) that introduces semantic priors on top of instance discrimination for visual UDA tasks. By considering instance contrastive learning as a dictionary look-up operation, we construct a semantics-aware dictionary with samples from both source and target domains where each target sample is assigned a (pseudo) category label based on the category priors of source samples. This allows category contrastive learning (between target queries and the category-level dictionary) for category-discriminative yet domain-invariant feature representations: samples of the same category (from either source or target domain) are pulled closer while those of different categories are pushed apart simultaneously. Extensive UDA experiments in multiple visual tasks (e.g., segmentation, classification and detection) show that CaCo achieves superior performance as compared with state-of-the-art methods. The experiments also demonstrate that CaCo is complementary to existing UDA methods and gen-eralizable to other learning setups such as unsupervised model adaptation, open-/partial-set adaptation etc.

Recent studies show that, both explicit deep feature matching as well as large-scale and diverse training data can significantly improve the generalization of person reidentification. However, the efficiency of learning deep matchers on large-scale data has not yet been adequately studied. Though learning with classification parameters or class memory is a popular way, it incurs large memory and computational costs. In contrast, pairwise deep metric learning within mini batches would be a better choice. However, the most popular random sampling method, the well-known PKsampler, is not informative and efficient for deep metric learning. Though online hard example mining has improved the learning efficiency to some extent, the mining in mini batches after random sampling is still limited. This inspires us to explore the use of hard example mining earlier, in the data sampling stage. To do so, in this paper, we propose an efficient mini-batch sampling method, called graph sampling (GS), for large-scale deep metric learning. The basic idea is to build a nearest neighbor relationship graph for all classes at the beginning of each epoch. Then, each mini batch is composed of a randomly selected class and its nearest neighboring classes so as to provide informative and challenging examples for learning. Together with an adapted competitive baseline, we improve the state of the art in generalizable person re-identification significantly, by 25.1% in Rank-1 on MSMT17 when trained on RandPerson. Besides, the proposed method also outperforms the competitive baseline, by 6.8% in Rank-1 on CUHK03-NP when trained on MSMT17. Meanwhile, the training time is significantly reduced, from 25.4 hours to 2 hours when trained on RandPerson with 8,000 identities. Code is available at https://github.com/ShengcaiLiao/QAConv.

Existing studies in few-shot semantic segmentation only focus on mining the target object information, however, often are hard to tell ambiguous regions, especially in non-target regions, which include background (BG) and Distracting Objects (DOs). To alleviate this problem, we propose a novel framework, namely Non-Target Region Eliminating (NTRE) network, to explicitly mine and eliminate BG and DO regions in the query. First, a BG Mining Module (BGMM) is proposed to extract the BG region via learning a general BG prototype. To this end, we design a BG loss to supervise the learning of BGMM only using the known target object segmentation ground truth. Then, a BG Eliminating Module and a DO Eliminating Module are proposed to successively filter out the BG and DO information from the query feature, based on which we can obtain a BG and DO-free target object segmentation result. Furthermore, we propose a prototypical contrastive learning algorithm to improve the model ability of distinguishing the target object from DOs. Extensive experiments on both PASCAL-5iand COCO-20idatasets show that our approach is effective despite its simplicity. Code is available at https://github.com/LIUYUANWEI98/NERTNet

Referring video object segmentation (RVOS) is a challenging language-guided video grounding task, which requires comprehensively understanding the semantic information of both video content and language queries for object prediction. However, existing methods adopt multi-modal fusion at a frame-based spatial granularity. The limitation of visual representation is prone to causing vision-language mismatching and producing poor segmentation results. To address this, we propose a novel multi-level representation learning approach, which explores the inherent structure of the video content to provide a set of discriminative visual embedding, enabling more effective vision-language semantic alignment. Specifically, we embed different visual cues in terms of visual granularity, including multi-frame long-temporal information at video level, intra-frame spatial semantics at frame level, and enhanced object-aware feature prior at object level. With the powerful multi-level visual embedding and carefully-designed dynamic alignment, our model can generate a robust representation for accurate video object segmentation. Extensive experiments on Refer-DAVIS17and Refer-YouTube-VOS demonstrate that our model achieves superior performance both in segmentation accuracy and inference speed.

This paper strives for activity recognition under domain shift, for example caused by change of scenery or camera viewpoint. The leading approaches reduce the shift in activity appearance by adversarial training and self-supervised learning. Different from these vision-focused works we leverage activity sounds for domain adaptation as they have less variance across domains and can reliably indicate which activities are not happening. We propose an audio-adaptive encoder and associated learning methods that discriminatively adjust the visual feature representation as well as addressing shifts in the semantic distribution. To further eliminate domain-specific features and include domain-invariant activity sounds for recognition, an audio-infused recognizer is proposed, which effectively models the cross-modal interaction across domains. We also introduce the new task of actor shift, with a corresponding audio-visual dataset, to challenge our method with situations where the activity appearance changes dramatically. Experiments on this dataset, EPIC-Kitchens and CharadesEgo show the effectiveness of our approach. Project page: https://xiaobai1217.github.io/DomainAdaptation.

This paper presents a Refinement Pyramid Transformer (RePFormer) for robust facial landmark detection. Most facial landmark detectors focus on learning representative image features. However, these CNN-based feature representations are not robust enough to handle complex real-world scenarios due to ignoring the internal structure of landmarks, as well as the relations between landmarks and context. In this work, we formulate the facial landmark detection task as refining landmark queries along pyramid memories. Specifically, a pyramid transformer head (PTH) is introduced to build both homologous relations among landmarks and heterologous relations between landmarks and cross-scale contexts. Besides, a dynamic landmark refinement (DLR) module is designed to decompose the landmark regression into an end-to-end refinement procedure, where the dynamically aggregated queries are transformed to residual coordinates predictions. Extensive experimental results on four facial landmark detection benchmarks and their various subsets demonstrate the superior performance and high robustness of our framework.

LiDAR point clouds, which are usually scanned by rotating LiDAR sensors continuously, capture precise geometry of the surrounding environment and are crucial to many autonomous detection and navigation tasks. Though many 3D deep architectures have been developed, efficient collection and annotation of large amounts of point clouds remain one major challenge in the analytics and understanding of point cloud data. This paper presents PolarMix, a point cloud augmentation technique that is simple and generic but can mitigate the data constraint effectively across various perception tasks and scenarios. PolarMix enriches point cloud distributions and preserves point cloud fidelity via two cross-scan augmentation strategies that cut, edit, and mix point clouds along the scanning direction. The first is scene-level swapping which exchanges point cloud sectors of two LiDAR scans that are cut along the LiDAR scanning direction. The second is instance-level rotation and paste which crops point instances from one LiDAR scan, rotates them by multiple angles (to create multiple copies), and paste the rotated point instances into other scans. Extensive experiments show that PolarMix achieves superior performance consistently across different perception tasks and scenarios. In addition, it can work as a plug-and-play for various 3D deep architectures and also performs well for unsupervised domain adaptation.

Transformers have recently lead to encouraging progress in computer vision. In this work, we present new baselines by improving the original Pyramid Vision Transformer (PVT v1) by adding three designs: (i) a linear complexity attention layer, (ii) an overlapping patch embedding, and (iii) a convolutional feed-forward network. With these modifications, PVT v2 reduces the computational complexity of PVT v1 to linearity and provides significant improvements on fundamental vision tasks such as classification, detection, and segmentation. In particular, PVT v2 achieves comparable or better performance than recent work such as the Swin transformer. We hope this work will facilitate state-of-the-art transformer research in computer vision. Code is available at https://github.com/whai362/PVT .

Automating report generation for medical imaging promises to minimize labor and aid diagnosis in clinical practice. Deep learning algorithms have recently been shown to be capable of captioning natural photos. However, doing a similar thing for medical data, is difficult due to the variety in reports written by different radiologists with fluctuating levels of knowledge and experience. Current methods for automatic report generation tend to merely copy one of the training samples in the created report. To tackle this issue, we propose variational topic inference, a probabilistic approach for automatic chest X-ray report generation. Specifically, we introduce a probabilistic latent variable model where a latent variable defines a single topic. The topics are inferred in a conditional variational inference framework by aligning vision and language modalities in a latent space, with each topic governing the generation of one sentence in the report. We further adopt a visual attention module that enables the model to attend to different locations in the image while generating the descriptions. We conduct extensive experiments on two benchmarks, namely Indiana U. Chest X-rays and MIMIC-CXR. The results demonstrate that our proposed variational topic inference method can generate reports with novel sentence structure, rather than mere copies of reports used in training, while still achieving comparable performance to state-of-the-art methods in terms of standard language generation criteria.

Pedestrian detection is an important but challenging problem in computer vision, especially in human-centric tasks. Over the past decade, significant improvement has been witnessed with the help of handcrafted features and deep features. Here we present a comprehensive survey on recent advances in pedestrian detection. First, we provide a detailed review of single-spectral pedestrian detection that includes handcrafted features based methods and deep features based approaches. For handcrafted features based methods, we present an extensive review of approaches and find that handcrafted features with large freedom degrees in shape and space have better performance. In the case of deep features based approaches, we split them into pure CNN based methods and those employing both handcrafted and CNN based features. We give the statistical analysis and tendency of these methods, where feature enhanced, part-aware, and post-processing methods have attracted main attention. In addition to single-spectral pedestrian detection, we also review multi-spectral pedestrian detection, which provides more robust features for illumination variance. Furthermore, we introduce some related datasets and evaluation metrics, and a deep experimental analysis. We conclude this survey by emphasizing open problems that need to be addressed and highlighting various future directions. Researchers can track an up-to-date list at https://github.com/JialeCao001/PedSurvey.

Generic object counting in natural scenes is a challenging computer vision problem. Existing approaches either rely on instance-level supervision or absolute count information to train a generic object counter. We introduce a partially supervised setting that significantly reduces the supervision level required for generic object counting. We propose two novel frameworks, named lower-count (LC) and reduced lower-count (RLC), to enable object counting under this setting. Our frameworks are built on a novel dual-branch architecture that has an image classification and a density branch. Our LC framework reduces the annotation cost due to multiple instances in an image by using only lower-count supervision for all object categories. Our RLC framework further reduces the annotation cost arising from large numbers of object categories in a dataset by only using lower-count supervision for a subset of categories and class-labels for the remaining ones. The RLC framework extends our dual-branch LC framework with a novel weight modulation layer and a category-independent density map prediction. Experiments are performed on COCO, Visual Genome and PASCAL 2007 datasets. Our frameworks perform on par with state-of-the-art approaches using higher levels of supervision. Additionally, we demonstrate the applicability of our LC supervised density map for image-level supervised instance segmentation.

We present the first systematic study on concealed object detection (COD), which aims to identify objects that are visually embedded in their background. The high intrinsic similarities between the concealed objects and their background make COD far more challenging than traditional object detection/segmentation. To better understand this task, we collect a large-scale dataset, called COD10K, which consists of 10,000 images covering concealed objects in diverse real-world scenarios from 78 object categories. Further, we provide rich annotations including object categories, object boundaries, challenging attributes, object-level labels, and instance-level annotations. Our COD10K is the largest COD dataset to date, with the richest annotations, which enables comprehensive concealed object understanding and can even be used to help progress several other vision tasks, such as detection, segmentation, classification etc. Motivated by how animals hunt in the wild, we also design a simple but strong baseline for COD, termed the Search Identification Network (SINet). Without any bells and whistles, SINet outperforms twelve cutting-edge baselines on all datasets tested, making them robust, general architectures that could serve as catalysts for future research in COD. Finally, we provide some interesting findings, and highlight several potential applications and future directions. To spark research in this new field, our code, dataset, and online demo are available at our project page: http://mmcheng.net/cod.

Conventional salient object detection models cannot differentiate the importance of different salient objects. Recently, two works have been proposed to detect saliency ranking by assigning different degrees of saliency to different objects. However, one of these models cannot differentiate object instances and the other focuses more on sequential attention shift order inference. In this paper, we investigate a practical problem setting that requires simultaneously segment salient instances and infer their relative saliency rank order. We present a novel unified model as the first end-to-end solution, where an improved Mask R-CNN is first used to segment salient instances and a saliency ranking branch is then added to infer the relative saliency. For relative saliency ranking, we build a new graph reasoning module by combining four graphs to incorporate the instance interaction relation, local contrast, global contrast, and a high-level semantic prior, respectively. A novel loss function is also proposed to effectively train the saliency ranking branch. Besides, a new dataset and an evaluation metric are proposed for this task, aiming at pushing forward this field of research. Finally, experimental results demonstrate that our proposed model is more effective than previous methods. We also show an example of its practical usage on adaptive image retargeting.

How to effectively fuse cross-modal information is a key problem for RGB-D salient object detection. Early fusion and result fusion schemes fuse RGB and depth information at the input and output stages, respectively, and hence incur distribution gaps or information loss. Many models instead employ a feature fusion strategy, but they are limited by their use of low-order point-to-point fusion methods. In this paper, we propose a novel mutual attention model by fusing attention and context from different modalities. We use the non-local attention of one modality to propagate long-range contextual dependencies for the other, thus leveraging complementary attention cues to achieve high-order and trilinear cross-modal interaction. We also propose to induce contrast inference from the mutual attention and obtain a unified model. Considering that low-quality depth data may be detrimental to model performance, we further propose a selective attention to reweight the added depth cues. We embed the proposed modules in a two-stream CNN for RGB-D SOD. Experimental results demonstrate the effectiveness of our proposed model. Moreover, we also construct a new and challenging large-scale RGB-D SOD dataset of high-quality, which can promote both the training and evaluation of deep models.

Binary optimization problems (BOPs) arise naturally in many fields, such as information retrieval, computer vision, and machine learning. Most existing binary optimization methods either use continuous relaxation which can cause large quantization errors, or incorporate a highly specific algorithm that can only be used for particular loss functions. To overcome these difficulties, we propose a novel generalized optimization method, named Alternating Binary Matrix Optimization (ABMO), for solving BOPs. ABMO can handle BOPs with/without orthogonality or linear constraints for a large class of loss functions. ABMO involves rewriting the binary, orthogonality and linear constraints for BOPs as an intersection of two closed sets, then iteratively dividing the original problems into several small optimization problems that can be solved as closed forms. To provide a strict theoretical convergence analysis, we add a sufficiently small perturbation and translate the original problem to an approximated problem whose feasible set is continuous. We not only provide rigorous mathematical proof for the convergence to a stationary and feasible point, but also derive the convergence rate of the proposed algorithm. The promising results obtained from four binary optimization tasks validate the superiority and the generality of ABMO compared with the state-of-the-art methods.

Person re-identification (Re-ID) aims at retrieving a person of interest across multiple non-overlapping cameras. With the advancement of deep neural networks and increasing demand of intelligent video surveillance, it has gained significantly increased interest in the computer vision community. By dissecting the involved components in developing a person Re-ID system, we categorize it into the closed-world and open-world settings. The widely studied closed-world setting is usually applied under various research-oriented assumptions, and has achieved inspiring success using deep learning techniques on a number of datasets. We first conduct a comprehensive overview with in-depth analysis for closed-world person Re-ID from three different perspectives, including deep feature representation learning, deep metric learning and ranking optimization. With the performance saturation under closed-world setting, the research focus for person Re-ID has recently shifted to the open-world setting, facing more challenging issues. This setting is closer to practical applications under specific scenarios. We summarize the open-world Re-ID in terms of five different aspects. By analyzing the advantages of existing methods, we design a powerful AGW baseline, achieving state-of-the-art or at least comparable performance on twelve datasets for four different Re-ID tasks. Meanwhile, we introduce a new evaluation metric (mINP) for person Re-ID, indicating the cost for finding all the correct matches, which provides an additional criteria to evaluate the Re-ID system for real applications. Finally, some important yet under-investigated open issues are discussed.

Existing domain adaptation (DA) methods generally assume that different domains have identical label space, and the training data are only sampled from a single domain. This unrealistic assumption is quite restricted for real-world applications, since it neglects the more practical scenario, where the source domain can contain the categories that are not shared by the target domain, and the training data can be collected from multiple modalities. In this article, we address a more difficult but practical problem, which recognizes RGB images through training on RGB-D data under the label space inequality scenario. There are three challenges in this task: 1) source and target domains are affected by the domain mismatch issue, which results in that the trained models perform imperfectly on the test data; 2) depth images are absent in the target domain (e.g., target images are captured by smartphones), when the source domain contains both the RGB and depth data. It makes the ordinary visual recognition approaches hardly applied to this task; and 3) in the real world, the source and target domains always have different numbers of categories, which would result in a negative transfer bottleneck being more prominent. Toward tackling the above challenges, we formulate a deep model, called visual-depth matching network (VDMN), where two new modules and a matching component can be trained in an end-to-end fashion jointly to identify the common and outlier categories effectively. The significance of VDMN is that it can take advantage of depth information and handle the domain distribution mismatch under label inequality simultaneously. The experimental results reveal that VDMN exceeds the state-of-the-art performance on various DA datasets, especially under the label inequality scenario.

State-of-the-art object detectors usually progressively downsample the input image until it is represented by small feature maps, which loses the spatial information and compromises the representation of small objects. In this article, we propose a context-aware block net (CAB Net) to improve small object detection by building high-resolution and strong semantic feature maps. To internally enhance the representation capacity of feature maps with high spatial resolution, we delicately design the context-aware block (CAB). CAB exploits pyramidal dilated convolutions to incorporate multilevel contextual information without losing the original resolution of feature maps. Then, we assemble CAB to the end of the truncated backbone network (e.g., VGG16) with a relatively small downsampling factor (e.g., 8) and cast off all following layers. CAB Net can capture both basic visual patterns as well as semantical information of small objects, thus improving the performance of small object detection. Experiments conducted on the benchmark Tsinghua-Tencent 100K and the Airport dataset show that CAB Net outperforms other top-performing detectors by a large margin while keeping real-time speed, which demonstrates the effectiveness of CAB Net for small object detection.

Computational photo quality evaluation is a useful technique in many tasks of computer vision and graphics, for example, photo retaregeting, 3-D rendering, and fashion recommendation. The conventional photo quality models are designed by characterizing the pictures from all communities (e.g., "architecture" and "colorful") indiscriminately, wherein community-specific features are not exploited explicitly. In this article, we develop a new community-aware photo quality evaluation framework. It uncovers the latent community-specific topics by a regularized latent topic model (LTM) and captures human visual quality perception by exploring multiple attributes. More specifically, given massive-scale online photographs from multiple communities, a novel ranking algorithm is proposed to measure the visual/semantic attractiveness of regions inside each photograph. Meanwhile, three attributes, namely: 1) photo quality scores; weak semantic tags; and inter-region correlations, are seamlessly and collaboratively incorporated during ranking. Subsequently, we construct the gaze shifting path (GSP) for each photograph by sequentially linking the top-ranking regions from each photograph, and an aggregation-based CNN calculates the deep representation for each GSP. Based on this, an LTM is proposed to model the GSP distribution from multiple communities in the latent space. To mitigate the overfitting problem caused by communities with very few photographs, a regularizer is incorporated into our LTM. Finally, given a test photograph, we obtain its deep GSP representation and its quality score is determined by the posterior probability of the regularized LTM. Comparative studies on four image sets have shown the competitiveness of our method. Besides, the eye-tracking experiments have demonstrated that our ranking-based GSPs are highly consistent with real human gaze movements.

Accurately classifying sceneries with different spatial configurations is an indispensable technique in computer vision and intelligent systems, for example, scene parsing, robot motion planning, and autonomous driving. Remarkable performance has been achieved by the deep recognition models in the past decade. As far as we know, however, these deep architectures are incapable of explicitly encoding the human visual perception, that is, the sequence of gaze movements and the subsequent cognitive processes. In this article, a biologically inspired deep model is proposed for scene classification, where the human gaze behaviors are robustly discovered and represented by a unified deep active learning (UDAL) framework. More specifically, to characterize objects' components with varied sizes, an objectness measure is employed to decompose each scenery into a set of semantically aware object patches. To represent each region at a low level, a local-global feature fusion scheme is developed which optimally integrates multimodal features by automatically calculating each feature's weight. To mimic the human visual perception of various sceneries, we develop the UDAL that hierarchically represents the human gaze behavior by recognizing semantically important regions within the scenery. Importantly, UDAL combines the semantically salient region detection and the deep gaze shifting path (GSP) representation learning into a principled framework, where only the partial semantic tags are required. Meanwhile, by incorporating the sparsity penalty, the contaminated/redundant low-level regional features can be intelligently avoided. Finally, the learned deep GSP features from the entire scene images are integrated to form an image kernel machine, which is subsequently fed into a kernel SVM to classify different sceneries. Experimental evaluations on six well-known scenery sets (including remote sensing images) have shown the competitiveness of our approach.

Matching the daytime visible and nighttime infrared person images, namely visible infrared person re-identification (VI-ReID), is a challenging cross-modality retrieval problem. Due to the difficulty of data collection and annotation in nighttime surveillance, VI-ReID usually suffers from noise problems, making it challenging to directly learn part discriminative features. In order to improve the discriminability and enhance the robustness against noisy images, this paper proposes a novel dynamic tri-level relation mining (DTRM) framework by simultaneously exploring channel-level, part-level intra-modality, and graph-level cross-modality relation cues. To address the misalignment within the person images, we design an intra-modality weighted-part attention (IWPA) to construct part-aggregated representation. It adaptively integrates the body part relation into the local feature learning with a residual batch normalization (RBN) connection scheme. Besides, a cross-modality graph structured attention (CGSA) is incorporated to improve the global feature learning by utilizing the contextual relation between images from two modalities. This module reduces the negative effects of noisy images. To seamlessly integrate two components, a parameter-free dynamic aggregation strategy is designed in a progressive joint learning manner. To further improve the performance, we additionally design a simple yet effective channel-level learning strategy by exploiting the rich channel information of visible images, which significantly reinforces the performance without modifying the network structure or changing the training process. Extensive experiments on two visible infrared re-identification datasets have verified the effectiveness under various settings. Code is available at:https://github.com/mangye16/DDAG

Spatial-temporal relation reasoning is a significant yet challenging problem for video action recognition. Previous works typically apply local operations like 2D or 3D CNNs to conduct space-time interactions in video sequences, or simply capture space-time long-range relations of a single fixed scale. However, this is inadequate for obtaining a comprehensive action representation. Besides, most models treat all input frames equally for the final classification, without selecting key frames and motion-sensitive regions. This introduces irrelevant video content and hurts the performance of models. In this paper, we propose a generic Spatial-Temporal Pyramid Graph Network (STPG-Net) to adaptively capture long-range spatial-temporal relations in video sequences at multiple scales. Specifically, we design a temporal attention (TA) module and a spatial-temporal attention (STA) module to learn the contribution of each frame and each space-time region to an action at a feature level, respectively. We then apply the selected key information to build spatial-temporal pyramid graphs for long-range relation reasoning and more comprehensive action representation learning. STPG-Net can be flexibly integrated into 2D and 3D backbone networks in a plug-and-play manner. Extensive experiments show that it brings consistent improvements over many challenging baselines on several standard action recognition benchmarks (i.e., Something-Something V1 & V2, and FineGym), demonstrating the effectiveness of our approach.

The success of current deep saliency models heavily depends on large amounts of annotated human fixation data to fit the highly non-linear mapping between the stimuli and visual saliency. Such fully supervised data-driven approaches are annotation-intensive and often fail to consider the underlying mechanisms of visual attention. In contrast, in this paper, we introduce a model based on various cognitive theories of visual saliency, which learns visual attention patterns in a weakly supervised manner. Our approach incorporates insights from cognitive science as differentiable submodules, resulting in a unified, end-to-end trainable framework. Specifically, our model encapsulates the following important components motivated from biological vision. (a) As scene semantics are closely related to visually attentive regions, our model encodes discriminative spatial information for scene understanding through spatial visual semantics embedding. (b) To model the objectness factors in visual attention deployment, we incorporate object-level semantics embedding and object relation information. (c) Considering the "winner-take-all" mechanism in visual stimuli processing, we model the competition mechanism among objects with softmax based neural attention. (d) Lastly, a conditional center prior is learned to mimic the spatial distribution bias of visual attention. Furthermore, we propose novel loss functions to utilize supervision cues from image-level semantics, saliency prior knowledge, and self-information compression. Experiments show that our method achieves promising results, and even outperforms many of its fully supervised counterparts. Overall, our weakly supervised saliency method makes an essential step towards reducing the annotation budget of current approaches, as well as providing a more comprehensive understanding of the visual attention mechanism. Our code is available at: https://github.com/ashleylqx/WeakFixation.git.

Abnormal crowd behavior detection has recently attracted increasing attention due to its wide applications in computer vision research areas. However, it is still an extremely challenging task due to the great variability of abnormal behavior coupled with huge ambiguity and uncertainty of video contents. To tackle these challenges, we propose a new probabilistic framework named variational abnormal behavior detection (VABD), which can detect abnormal crowd behavior in video sequences. We make three major contributions: (1) We develop a new probabilistic latent variable model that combines the strengths of the U-Net and conditional variational auto-encoder, which also are the backbone of our model; (2) We propose a motion loss based on an optical flow network to impose the motion consistency of generated video frames and input video frames; (3) We embed a Wasserstein generative adversarial network at the end of the backbone network to enhance the framework performance. VABD can accurately discriminate abnormal video frames from video sequences. Experimental results on UCSD, CUHK Avenue, IITB-Corridor, and ShanghaiTech datasets show that VABD outperforms the state-of-the-art algorithms on abnormal crowd behavior detection. Without data augmentation, our VABD achieves 72.24% in terms of AUC on IITB-Corridor, which surpasses the state-of-the-art methods by nearly 5%.

Unsupervised domain adaptation (UDA) for person re-identification is challenging because of the huge gap between the source and target domain. A typical self-training method is to use pseudo-labels generated by clustering algorithms to iteratively optimize the model on the target domain. However, a drawback to this is that noisy pseudo-labels generally cause trouble in learning. To address this problem, a mutual learning method by dual networks has been developed to produce reliable soft labels. However, as the two neural networks gradually converge, their complementarity is weakened and they likely become biased towards the same kind of noise. This paper proposes a novel light-weight module, the Attentive WaveBlock (AWB), which can be integrated into the dual networks of mutual learning to enhance the complementarity and further depress noise in the pseudo-labels. Specifically, we first introduce a parameter-free module, the WaveBlock, which creates a difference between features learned by two networks by waving blocks of feature maps differently. Then, an attention mechanism is leveraged to enlarge the difference created and discover more complementary features. Furthermore, two kinds of combination strategies, i.e. pre-attention and post-attention, are explored. Experiments demonstrate that the proposed method achieves state-of-the-art performance with significant improvements on multiple UDA person re-identification tasks. We also prove the generality of the proposed method by applying it to vehicle re-identification and image classification tasks. Our codes and models are available at: AWB.

Existing person re-identification (Re-ID) methods usually rely heavily on large-scale thoroughly annotated training data. However, label noise is unavoidable due to inaccurate person detection results or annotation errors in real scenes. It is extremely challenging to learn a robust Re-ID model with label noise since each identity has very limited annotated training samples. To avoid fitting to the noisy labels, we propose to learn a prefatory model using a large learning rate at the early stage with a self-label refining strategy, in which the labels and network are jointly optimized. To further enhance the robustness, we introduce an online co-refining (CORE) framework with dynamic mutual learning, where networks and label predictions are online optimized collaboratively by distilling the knowledge from other peer networks. Moreover, it also reduces the negative impact of noisy labels using a favorable selective consistency strategy. CORE has two primary advantages: it is robust to different noise types and unknown noise ratios; it can be easily trained without much additional effort on the architecture design. Extensive experiments on Re-ID and image classification demonstrate that CORE outperforms its counterparts by a large margin under both practical and simulated noise settings. Notably, it also improves the state-of-the-art unsupervised Re-ID performance under standard settings. Code is available at https://github.com/mangye16/ReID-Label-Noise.

Inferring the scene illumination from a single image is an essential yet challenging task in computer vision and computer graphics. Existing works estimate lighting by regressing representative illumination parameters or generating illumination maps directly. However, these methods often suffer from poor accuracy and generalization. This paper presents Geometric Mover's Light (GMLight), a lighting estimation framework that employs a regression network and a generative projector for effective illumination estimation. We parameterize illumination scenes in terms of the geometric light distribution, light intensity, ambient term, and auxiliary depth, which can be estimated by a regression network. Inspired by the earth mover's distance, we design a novel geometric mover's loss to guide the accurate regression of light distribution parameters. With the estimated light parameters, the generative projector synthesizes panoramic illumination maps with realistic appearance and high-frequency details. Extensive experiments show that GMLight achieves accurate illumination estimation and superior fidelity in relighting for 3D object insertion. The codes are available at https://github.com/fnzhan/Illumination-Estimation.

Acquiring sufficient ground-truth supervision to train deep visual models has been a bottleneck over the years due to the data-hungry nature of deep learning. This is exacerbated in some structured prediction tasks, such as semantic segmentation, which require pixel-level annotations. This work addresses weakly supervised semantic segmentation (WSSS), with the goal of bridging the gap between image-level annotations and pixel-level segmentation. To achieve this, we propose, for the first time, a novel group-wise learning framework for WSSS. The framework explicitly encodes semantic dependencies in a group of images to discover rich semantic context for estimating more reliable pseudo ground-truths, which are subsequently employed to train more effective segmentation models. In particular, we solve the group-wise learning within a graph neural network (GNN), wherein input images are represented as graph nodes, and the underlying relations between a pair of images are characterized by graph edges. We then formulate semantic mining as an iterative reasoning process which propagates the common semantics shared by a group of images to enrich node representations. Moreover, in order to prevent the model from paying excessive attention to common semantics, we further propose a graph dropout layer to encourage the graph model to capture more accurate and complete object responses. With the above efforts, our model lays the foundation for more sophisticated and flexible group-wise semantic mining. We conduct comprehensive experiments on the popular PASCAL VOC 2012 and COCO benchmarks, and our model yields state-of-the-art performance. In addition, our model shows promising performance in weakly supervised object localization (WSOL) on the CUB-200-2011 dataset, demonstrating strong generalizability. Our code is available at: https://github.com/Lixy1997/Group-WSSS.

Diabetic retinopathy (DR) is a complication of diabetes that severely affects eyes. It can be graded into five levels of severity according to international protocol. However, optimizing a grading model to have strong generalizability requires a large amount of balanced training data, which is difficult to collect, particularly for the high severity levels. Typical data augmentation methods, including random flipping and rotation, cannot generate data with high diversity. In this paper, we propose a diabetic retinopathy generative adversarial network (DR-GAN) to synthesize high-resolution fundus images which can be manipulated with arbitrary grading and lesion information. Thus, large-scale generated data can be used for more meaningful augmentation to train a DR grading and lesion segmentation model. The proposed retina generator is conditioned on the structural and lesion masks, as well as adaptive grading vectors sampled from the latent grading space, which can be adopted to control the synthesized grading severity. Moreover, a multi-scale spatial and channel attention module is devised to improve the generation ability to synthesize small details. Multi-scale discriminators are designed to operate from large to small receptive fields, and joint adversarial losses are adopted to optimize the whole network in an end-to-end manner. With extensive experiments evaluated on the EyePACS dataset connected to Kaggle, as well as the FGADR dataset, we validate the effectiveness of our method, which can both synthesize highly realistic ( 1280 ×1280) controllable fundus images and contribute to the DR grading task.

Text segmentation is a fundamental task in natural language processing. Depending on the levels of granularity, the task can be defined as segmenting a document into topical segments, or segmenting a sentence into elementary discourse units (EDUs). Traditional solutions to the two tasks heavily rely on carefully designed features. The recently proposed neural models do not need manual feature engineering, but they either suffer from sparse boundary tags or cannot efficiently handle the issue of variable size output vocabulary. In light of such limitations, we propose a generic end-to-end segmentation model, namely${\mathrm{S}\scriptstyle{\mathrm{EG}}}{\mathrm{B}\scriptstyle{\mathrm{OT}}}$, which first uses a bidirectional recurrent neural network to encode an input text sequence.${\mathrm{S}\scriptstyle{\mathrm{EG}}}{\mathrm{B}\scriptstyle{\mathrm{OT}}}$then uses another recurrent neural networks, together with a pointer network, to select text boundaries in the input sequence. In this way,${\mathrm{S}\scriptstyle{\mathrm{EG}}}{\mathrm{B}\scriptstyle{\mathrm{OT}}}$does not require any hand-crafted features. More importantly,${\mathrm{S}\scriptstyle{\mathrm{EG}}}{\mathrm{B}\scriptstyle{\mathrm{OT}}}$inherently handles the issue of variable size output vocabulary and the issue of sparse boundary tags. In our experiments,${\mathrm{S}\scriptstyle{\mathrm{EG}}}{\mathrm{B}\scriptstyle{\mathrm{OT}}}$outperforms state-of-the-art models on two tasks: document-level topic segmentation and sentence-level EDU segmentation. As a downstream application, we further propose a hierarchical attention model for sentence-level sentiment analysis based on the outcomes of${\mathrm{S}\scriptstyle{\mathrm{EG}}}{\mathrm{B}\scriptstyle{\mathrm{OT}}}$. The hierarchical model can make full use of both word-level and EDU-level information simultaneously for sentence-level sentiment analysis. In particular, it can effectively exploit EDU-level information, such as the inner properties of EDUs, which cannot be fully encoded in word-level features. Experimental results show that our hierarchical model achieves new state-of-the-art results on the Movie Review and Stanford Sentiment Treebank benchmarks.

In the field of computer vision, fine-grained visual categorization has attracted a lot of attention and made great progress due to convolutional neural networks and a large number of publicly available datasets. With next-generation sensing technology, RGB-D cameras can provide high-quality synchronized RGB and depth images for solving many computer vision problems. Although RGB-D cameras have been used in the context of multi-view object category detection and scene understanding, they have not been widely used in fine-grained classification. In this paper, we introduce a multiview RGB-D dataset RGBD-FG for fine-grained categorization. Currently, the dataset contains 93 051 RGB-D images covering 19 super-categories and 50 sub-categories of common vegetables and fruit, and is organized in a hierarchical manner. We provide extensive experimental results to establish state-of-the-art benchmarks for our dataset, illustrating its diversity and scope for improvement through future work. We also propose a novel modality-specific multimodal network called FS-Multimodal network, which can solve two limitations of multimodal networks trained based on fine-tuning techniques: over-fitting and lack of effective depth-specific features. We hope that our study lays the foundations for fine-grained categorization of RGB-D data.

Generative adversarial networks (GANs) for (generalized) zero-shot learning (ZSL) aim to generate unseen image features when conditioned on unseen class embeddings, each of which corresponds to one unique category. Most existing works on GANs for ZSL generate features by merely feeding the seen image feature/class embedding (combined with random Gaussian noise) pairs into the generator/discriminator for a two-player minimax game. However, the structure consistency of the distributions among the real/fake image features, which may shift the generated features away from their real distribution to some extent, is seldom considered. In this paper, to align the weights of the generator for better structure consistency between real/fake features, we propose a novel multigraph adaptive GAN (MGA-GAN). Specifically, a Wasserstein GAN equipped with a classification loss is trained to generate discriminative features with structure consistency. MGA-GAN leverages the multigraph similarity structures between sliced seen real/fake feature samples to assist in updating the generator weights in the local feature manifold. Moreover, correlation graphs for the whole real/fake features are adopted to guarantee structure correlation in the global feature manifold. Extensive evaluations on four benchmarks demonstrate well the superiority of MGA-GAN over its state-of-the-art counterparts.

Categorizing aerial photographs with varied weather/lighting conditions and sophisticated geomorphic factors is a key module in autonomous navigation, environmental evaluation, and so on. Previous image recognizers cannot fulfill this task due to three challenges: 1) localizing visually/semantically salient regions within each aerial photograph in a weakly annotated context due to the unaffordable human resources required for pixel-level annotation; 2) aerial photographs are generally with multiple informative attributes (e.g., clarity and reflectivity), and we have to encode them for better aerial photograph modeling; and 3) designing a cross-domain knowledge transferal module to enhance aerial photograph perception since multiresolution aerial photographs are taken asynchronistically and are mutually complementary. To handle the above problems, we propose to optimize aerial photograph's feature learning by leveraging the low-resolution spatial composition to enhance the deep learning of perceptual features with a high resolution. More specifically, we first extract many BING-based object patches (Cheng et al., 2014) from each aerial photograph. A weakly supervised ranking algorithm selects a few semantically salient ones by seamlessly incorporating multiple aerial photograph attributes. Toward an interpretable aerial photograph recognizer indicative to human visual perception, we construct a gaze shifting path (GSP) by linking the top-ranking object patches and, subsequently, derive the deep GSP feature. Finally, a cross-domain multilabel SVM is formulated to categorize each aerial photograph. It leverages the global feature from low-resolution counterparts to optimize the deep GSP feature from a high-resolution aerial photograph. Comparative results on our compiled million-scale aerial photograph set have demonstrated the competitiveness of our approach. Besides, the eye-tracking experiment has shown that our ranking-based GSPs are over 92% consistent with the real human gaze shifting sequences.

Few-shot human-object interaction (FS-HOI) recognition aims at inferring new interactions between human actions and surrounding objects merely with a few available instances. It is beneficial to alleviate the long-tail and combinatorial explosion problems in human-object interaction (HOI). Nevertheless, the existing FS-HOI methods only focus on modeling the relationships between labeled samples and unlabeled samples in the Euclidean domain, which neglects the rich relational structures of the visual information among labeled samples and between human actions and objects. Accordingly, we tackle the few-shot HOI task in the non-Euclidean domain and present a graph-based model, namely, task-oriented high-order context graph network (THCG-Net). It contains a task attention module (TA-Module) and a high-order context graph module (HG-Module). In TA-Module, an attention mechanism is designed by utilizing task information to build a task-oriented space, in which the discriminative information for the current task (episode) is captured by embedding the visual features into the task-oriented space. The HG-Module is proposed to construct a task-level graph and takes the context information as high-order knowledge, which provides discriminative guidance for propagating visual information. It captures the discriminability among different categories while highlights the commonality of related categories adaptively, which effectively transfers knowledge to related categories. Extensive experimental results on two benchmark datasets, HICO-FS and TUHOI-FS, are provided. It demonstrates that our THCG-Net significantly outperforms the state-of-the-art approaches, which proves its impressive effectiveness in recognizing various human actions and surrounding objects in few-shot scenarios.

Magnetic resonance (MR) image acquisition is an inherently prolonged process, whose acceleration by obtaining multiple undersampled images simultaneously through parallel imaging has always been the subject of research. In this paper, we propose the Dual-Octave Convolution (Dual-OctConv), which is capable of learning multi-scale spatial-frequency features from both real and imaginary components, for fast parallel MR image reconstruction. By reformulating the complex operations using octave convolutions, our model shows a strong ability to capture richer representations of MR images, while at the same time greatly reducing the spatial redundancy. More specifically, the input feature maps and convolutional kernels are first split into two components (i.e., real and imaginary), which are then divided into four groups according to their spatial frequencies. Then, our Dual-OctConv conducts intra-group information updating and inter-group information exchange to aggregate the contextual information across different groups. Our framework provides two appealing benefits: (i) it encourages interactions between real and imaginary components at various spatial frequencies to achieve richer representational capacity, and (ii) it enlarges the receptive field by learning multiple spatial-frequency features of both the real and imaginary components. We evaluate the performance of the proposed model on the acceleration of multi-coil MR image reconstruction. Extensive experiments are conducted on an {in vivo} knee dataset under different undersampling patterns and acceleration factors. The experimental results demonstrate the superiority of our model in accelerated parallel MR image reconstruction. Our code is available at: github.com/chunmeifeng/Dual-OctConv.

In this paper, we propose a domain-general model, termed learning-to-weight (LTW), that guarantees face detection performance across multiple domains, particularly the target domains that are never seen before. However, various face forgery methods cause complex and biased data distributions, making it challenging to detect fake faces in unseen domains. We argue that different faces contribute differently to a detection model trained on multiple domains, making the model likely to fit domain-specific biases. As such, we propose the LTW approach based on the meta-weight learning algorithm, which configures different weights for face images from different domains. The LTW network can balance the model's generalizability across multiple domains. Then, the meta-optimization calibrates the source domain's gradient enabling more discriminative features to be learned. The detection ability of the network is further improved by introducing an intra-class compact loss. Extensive experiments on several commonly used deepfake datasets to demonstrate the effectiveness of our method in detecting synthetic faces. Code and supplemental material are available at https://github.com/skJack/LTW.

Chest X-rays are an important and accessible clinical imaging tool for the detection of many thoracic diseases. Over the past decade, deep learning, with a focus on the convolutional neural network (CNN), has become the most powerful computer-aided diagnosis technology for improving disease identification performance. However, training an effective and robust deep CNN usually requires a large amount of data with high annotation quality. For chest X-ray imaging, annotating large-scale data requires professional domain knowledge and is time-consuming. Thus, existing public chest X-ray datasets usually adopt language pattern based methods to automatically mine labels from reports. However, this results in label uncertainty and inconsistency. In this paper, we propose many-to-one distribution learning (MODL) and K-nearest neighbor smoothing (KNNS) methods from two perspectives to improve a single model's disease identification performance, rather than focusing on an ensemble of models. MODL integrates multiple models to obtain a soft label distribution for optimizing the single target model, which can reduce the effects of original label uncertainty. Moreover, KNNS aims to enhance the robustness of the target model to provide consistent predictions on images with similar medical findings. Extensive experiments on the public NIH Chest X-ray and CheXpert datasets show that our model achieves consistent improvements over the state-of-the-art methods.

Batch normalization (BN) is an important technique commonly incorporated into deep learning models to perform standardization within mini-batches. The merits of BN in improving a model’s learning efficiency can be further amplified by applying whitening, while its drawbacks in estimating population statistics for inference can be avoided through group normalization (GN). This paper proposes group whitening (GW), which exploits the advantages of the whitening operation and avoids the disadvantages of normalization within mini-batches. In addition, we analyze the constraints imposed on features by normalization, and show how the batch size (group number) affects the performance of batch (group) normalized networks, from the perspective of model’s representational capacity. This analysis provides theoretical guidance for applying GW in practice. Finally, we apply the proposed GW to ResNet and ResNeXt architectures and conduct experiments on the ImageNet and COCO benchmarks. Results show that GW consistently improves the performance of different architectures, with absolute gains of 1.02% ∼ 1.49% in top-1 accuracy on ImageNet and 1.82% ∼ 3.21% in bounding box AP on COCO.

We present a novel group collaborative learning framework (GCoNet) capable of detecting co-salient objects in real time (16ms), by simultaneously mining consensus representations at group level based on the two necessary criteria: 1) intra-group compactness to better formulate the consistency among co-salient objects by capturing their inherent shared attributes using our novel group affinity module; 2) inter-group separability to effectively suppress the influence of noisy objects on the output by introducing our new group collaborating module conditioning the inconsistent consensus. To learn a better embedding space without extra computational overhead, we explicitly employ auxiliary classification supervision. Extensive experiments on three challenging benchmarks, i.e., CoCA, CoSOD3k, and Cosal2015, demonstrate that our simple GCoNet outperforms 10 cutting-edge models and achieves the new state-of-the-art. We demonstrate this paper’s new technical contributions on a number of important downstream computer vision applications including content aware co-segmentation, co-localization based automatic thumbnails, etc. Code has been made publicly available: https://github.com/fanq15/GCoNet.

Recently, Siamese-based trackers have achieved promising performance in visual tracking. Most recent Siamese-based trackers typically employ a depth-wise cross-correlation (DW-XCorr) to obtain multi-channel correlation information from the two feature maps (target and search region). However, DW-XCorr has several limitations within Siamese-based tracking: it can easily be fooled by distractors, has fewer activated channels and provides weak discrimination of object boundaries. Further, DW-XCorr is a handcrafted parameter-free module and cannot fully benefit from offline learning on large-scale data.We propose a learnable module, called the asymmetric convolution (ACM), which learns to better capture the se-mantic correlation information in offline training on large-scale data. Different from DW-XCorr and its predecessor (XCorr), which regard a single feature map as the convolution kernel, our ACM decomposes the convolution operation on a concatenated feature map into two mathematically equivalent operations, thereby avoiding the need for the feature maps to be of the same size (width and height) during concatenation. Our ACM can incorporate useful prior information, such as bounding-box size, with standard visual features. Furthermore, ACM can easily be integrated into existing Siamese trackers based on DW-XCorr or XCorr. To demonstrate its generalization ability, we integrate ACM into three representative trackers: SiamFC, SiamRPN++ and SiamBAN. Our experiments reveal the benefits of the proposed ACM, which outperforms existing methods on six tracking benchmarks. On the LaSOT test set, our ACM-based tracker obtains a significant improvement of 5.8% in terms of success (AUC), over the baseline.

Pedestrian detection is used in many vision based applications ranging from video surveillance to autonomous driving. Despite achieving high performance, it is still largely unknown how well existing detectors generalize to unseen data. This is important because a practical detector should be ready to use in various scenarios in applications. To this end, we conduct a comprehensive study in this paper, using a general principle of direct cross-dataset evaluation. Through this study, we find that existing state-of-the-art pedestrian detectors, though perform quite well when trained and tested on the same dataset, generalize poorly in cross dataset evaluation. We demonstrate that there are two reasons for this trend. Firstly, their designs (e.g. anchor settings) may be biased towards popular benchmarks in the traditional single-dataset training and test pipeline, but as a result largely limit their generalization capability. Secondly, the training source is generally not dense in pedestrians and diverse in scenarios. Under direct cross-dataset evaluation, surprisingly, we find that a general purpose object detector, without pedestrian-tailored adaptation in design, generalizes much better compared to existing state-of-the-art pedestrian detectors. Furthermore, we illustrate that diverse and dense datasets, collected by crawling the web, serve to be an efficient source of pre-training for pedestrian detection. Accordingly, we propose a progressive training pipeline and find that it works well for autonomous-driving oriented pedestrian detection. Consequently, the study conducted in this paper suggests that more emphasis should be put on cross-dataset evaluation for the future design of generalizable pedestrian detectors. Code and models can be accessed at https://github.com/hasanirtiza/Pedestron.

Recent advances in neuroscience have highlighted the effectiveness of multi-modal medical data for investigating certain pathologies and understanding human cognition. However, obtaining full sets of different modalities is limited by various factors, such as long acquisition times, high examination costs and artifact suppression. In addition, the complexity, high dimensionality and heterogeneity of neuroimaging data remains another key challenge in leveraging existing randomized scans effectively, as data of the same modality is often measured differently by different machines. There is a clear need to go beyond the traditional imaging-dependent process and synthesize anatomically specific target-modality data from a source in-put. In this paper, we propose to learn dedicated features that cross both intre- and intra-modal variations using a novel CSCℓ4Net. Through an initial unification of intra-modal data in the feature maps and multivariate canonical adaptation, CSC ℓ4Net facilitates feature-level mutual transformation. The positive definite Riemannian manifold-penalized data fidelity term further enables CSCℓ4Net to re-construct missing measurements according to transformed features. Finally, the maximization ℓ4-norm boils down to a computationally efficient optimization problem. Extensive experiments validate the ability and robustness of our CSC ℓ4Net compared to the state-of-the-art methods on multiple datasets.

In this paper, we propose a Monocular 3D Single Stage object Detector (M3DSSD) with feature alignment and asymmetric non-local attention. Current anchor-based monocular 3D object detection methods suffer from feature mismatching. To overcome this, we propose a two-step feature alignment approach. In the first step, the shape alignment is performed to enable the receptive field of the feature map to focus on the pre-defined anchors with high confidence scores. In the second step, the center alignment is used to align the features at 2D/3D centers. Further, it is often difficult to learn global information and capture long-range relationships, which are important for the depth prediction of objects. Therefore, we propose a novel asymmetric non-local attention block with multi-scale sampling to extract depth-wise features. The proposed M3DSSD achieves significantly better performance than the monocular 3D object detection methods on the KITTI dataset, in both 3D object detection and bird’s eye view tasks. The code is released at https://github.com/mumianyuxin/M3DSSD.

Few-shot semantic segmentation (FSS) aims to segment unseen class objects given very few densely-annotated support images from the same class. Existing FSS methods find the query object by using support prototypes or by directly relying on heuristic multi-scale feature fusion. However, they fail to fully leverage the high-order appearance relationships between multi-scale features among the support-query image pairs, thus leading to an inaccurate localization of the query objects. To tackle the above challenge, we propose an end-to-end scale-aware graph neural network (SAGNN) by reasoning the cross-scale relations among the support-query images for FSS. Specifically, a scale-aware graph is first built by taking support-induced multi-scale query features as nodes and, meanwhile, each edge is modeled as the pairwise interaction of its connected nodes. By progressive message passing over this graph, SAGNN is capable of capturing cross-scale relations and overcoming object variations (e.g., appearance, scale and location), and can thus learn more precise node embeddings. This in turn enables it to predict more accurate foreground objects. Moreover, to make full use of the location relations across scales for the query image, a novel self-node collaboration mechanism is proposed to enrich the current node, which endows SAGNN the ability of perceiving different resolutions of the same objects. Extensive experiments on PASCAL-5iand COCO-20ishow that SAGNN achieves state-of-the-art results.

Person search aims to simultaneously localize and identify a query person from realistic, uncropped images, which can be regarded as the unified task of pedestrian detection and person re-identification (re-id). Most existing works employ two-stage detectors like Faster-RCNN, yielding encouraging accuracy but with high computational overhead. In this work, we present the Feature-Aligned Person Search Network (AlignPS), the first anchor-free framework to efficiently tackle this challenging task. AlignPS explicitly addresses the major challenges, which we summarize as the misalignment issues in different levels (i.e., scale, region, and task), when accommodating an anchor-free detector for this task. More specifically, we propose an aligned feature aggregation module to generate more discriminative and robust feature embeddings by following a "re-id first" principle. Such a simple design directly improves the baseline anchor-free model on CUHK-SYSU by more than 20% in mAP. Moreover, AlignPS outperforms state-of-the-art two-stage methods, with a higher speed. The code is available at https://github.com/daodaofr/AlignPS.

Image restoration tasks demand a complex balance between spatial details and high-level contextualized information while recovering images. In this paper, we propose a novel synergistic design that can optimally balance these competing goals. Our main proposal is a multi-stage architecture, that progressively learns restoration functions for the degraded inputs, thereby breaking down the overall recovery process into more manageable steps. Specifically, our model first learns the contextualized features using encoder-decoder architectures and later combines them with a high-resolution branch that retains local information. At each stage, we introduce a novel per-pixel adaptive design that leverages in-situ supervised attention to reweight the local features. A key ingredient in such a multi-stage architecture is the information exchange between different stages. To this end, we propose a two-faceted approach where the information is not only exchanged sequentially from early to late stages, but lateral connections between feature processing blocks also exist to avoid any loss of information. The resulting tightly interlinked multi-stage architecture, named as MPRNet, delivers strong performance gains on ten datasets across a range of tasks including image deraining, deblurring, and denoising. The source code and pre-trained models are available at https://github.com/swz30/MPRNet.

This paper strives for repetitive activity counting in videos. Different from existing works, which all analyze the visual video content only, we incorporate for the first time the corresponding sound into the repetition counting process. This benefits accuracy in challenging vision conditions such as occlusion, dramatic camera view changes, low resolution, etc. We propose a model that starts with analyzing the sight and sound streams separately. Then an audiovisual temporal stride decision module and a reliability estimation module are introduced to exploit cross-modal temporal interaction. For learning and evaluation, an existing dataset is repurposed and reorganized to allow for repetition counting with sight and sound. We also introduce a variant of this dataset for repetition counting under challenging vision conditions. Experiments demonstrate the benefit of sound, as well as the other introduced modules, for repetition counting. Our sight-only model already outperforms the state-of-the-art by itself, when we add sound, results improve notably, especially under harsh vision conditions. The code and datasets are available at https://github.com/xiaobai1217/RepetitionCounting.

This paper addresses the task of unsupervised video multi-object segmentation. Current approaches follow a two-stage paradigm: 1) detect object proposals using pre-trained Mask R-CNN, and 2) conduct generic feature matching for temporal association using re-identification techniques. However, the generic features, widely used in both stages, are not reliable for characterizing unseen objects, leading to poor generalization. To address this, we introduce a novel approach for more accurate and efficient spatio-temporal segmentation. In particular, to address instance discrimination, we propose to combine foreground region estimation and instance grouping together in one network, and additionally introduce temporal guidance for segmenting each frame, enabling more accurate object discovery. For temporal association, we complement current video object segmentation architectures with a discriminative appearance model, capable of capturing more fine-grained target-specific information. Given object proposals from the instance discrimination network, three essential strategies are adopted to achieve accurate segmentation: 1) target-specific tracking using a memory-augmented appearance model; 2) target-agnostic verification to trace possible tracklets for the proposal; 3) adaptive memory updating using the verified segments. We evaluate the proposed approach on DAVIS17and YouTube-VIS, and the results demonstrate that it outperforms state-of-the-art methods both in segmentation accuracy and inference speed.

We present a new vision-language (VL) pre-training model dubbed Kaleido-BERT , which introduces a novel kaleido strategy for fashion cross-modality representations from transformers. In contrast to random masking strategy of recent VL models, we design alignment guided masking to jointly focus more on image-text semantic relations. To this end, we carry out five novel tasks, i.e., rotation, jigsaw, camouflage, grey-to-color, and blank-to-color for self-supervised VL pre-training at patches of different scale. Kaleido-BERT is conceptually simple and easy to extend to the existing BERT framework, it attains state-of-the-art results by large margins on four downstream tasks, including text retrieval (R@1: 4.03% absolute improvement), image retrieval (R@1: 7.13% abs imv.), category recognition (ACC: 3.28% abs imv.), and fashion captioning (Bleu4: 1.2 abs imv.). We validate the efficiency of Kaleido-BERT on a wide range of e-commerical websites, demonstrating its broader potential in real-world applications.

Generalized zero-shot learning (GZSL) has achieved significant progress, with many efforts dedicated to over-coming the problems of visual-semantic domain gap and seen-unseen bias. However, most existing methods directly use feature extraction models trained on ImageNet alone, ignoring the cross-dataset bias between ImageNet and GZSL benchmarks. Such a bias inevitably results in poor-quality visual features for GZSL tasks, which potentially limits the recognition performance on both seen and unseen classes. In this paper, we propose a simple yet effective GZSL method, termed feature refinement for generalized zero-shot learning (FREE), to tackle the above problem. FREE employs a feature refinement (FR) module that in-corporates semantic→visual mapping into a unified generative model to refine the visual features of seen and unseen class samples. Furthermore, we propose a self-adaptive margin center loss (SAMC-loss) that cooperates with a semantic cycle-consistency loss to guide FR to learn class- and semantically-relevant representations, and concatenate the features in FR to extract the fully refined features. Extensive experiments on five benchmark datasets demonstrate the significant performance gain of FREE over its baseline and current state-of-the-art methods. The code is available at https://github.com/shiming-chen/FREE.

While single-view 3D reconstruction has made significant progress benefiting from deep shape representations in recent years, garment reconstruction is still not solved well due to open surfaces, diverse topologies and complex geometric details. In this paper, we propose a novel learn-able Anchored Unsigned Distance Function (AnchorUDF) representation for 3D garment reconstruction from a single image. AnchorUDF represents 3D shapes by predicting unsigned distance fields (UDFs) to enable open garment surface modeling at arbitrary resolution. To capture diverse garment topologies, AnchorUDF not only computes pixel-aligned local image features of query points, but also leverages a set of anchor points located around the surface to enrich 3D position features for query points, which provides stronger 3D space context for the distance function. Furthermore, in order to obtain more accurate point projection direction at inference, we explicitly align the spatial gradient direction of AnchorUDF with the ground-truth direction to the surface during training. Extensive experiments on two public 3D garment datasets, i.e., MGN and Deep Fashion3D, demonstrate that AnchorUDF achieves the state-of-the-art performance on single-view garment reconstruction. Code is available at https://github.com/zhaofang0627/AnchorUDF.

A medical imaging model is usually explored for the diagnosis of a single disease. However, with the expanding demand for multi-disease diagnosis in clinical applications, multi-function solutions need to be investigated. Previous works proposed to either exploit different disease labels to conduct transfer learning through fine-tuning, or transfer knowledge across different domains with similar diseases. However, these methods still cannot address the real clinical challenge - a multi-disease model is required but annotations for each disease are not always available. In this paper, we introduce the task of transferring knowledge from single-disease diagnosis (source domain) to enhance multi-disease diagnosis (target domain). A category-invariant cross-domain transfer (CCT) method is proposed to address this single-to-multiple extension. First, for domain-specific task learning, we present a confidence weighted pooling (CWP) to obtain coarse heatmaps for different disease categories. Then, conditioned on these heatmaps, category-invariant feature refinement (CIFR) blocks are proposed to better localize discriminative semantic regions related to the corresponding diseases. The category-invariant characteristic enables transferability from the source domain to the target domain. We validate our method in two popular areas: extending diabetic retinopathy to identifying multiple ocular diseases, and extending glioma identification to the diagnosis of other brain tumors.

Automatic report generation on medical radiographs have recently gained interest. However, identifying diseases as well as correctly predicting their corresponding sizes, locations and other medical description patterns, which is essential for generating high-quality reports, is challenging. Although previous methods focused on producing readable reports, how to accurately detect and describe findings that match with the query X-Ray has not been successfully addressed. In this paper, we propose a multi-modality semantic attention model to integrate visual features, predicted key finding embeddings, as well as clinical features, and progressively decode reports with visual-textual semantic consistency. First, multi-modality features are extracted and attended with the hidden states from the sentence de-coder, to encode enriched context vectors for better decoding a report. These modalities include regional visual features of scans, semantic word embeddings of the top-K findings predicted with high probabilities, and clinical features of indications. Second, the progressive report decoder consists of a sentence decoder and a word decoder, where we propose image-sentence matching and description accuracy losses to constrain the visual-textual semantic consistency. Extensive experiments on the public MIMIC-CXR and IU X-Ray datasets show that our model achieves consistent improvements over the state-of-the-art methods.

Understanding the inner workings of deep neural networks (DNNs) is essential to provide trustworthy artificial intelligence techniques for practical applications. Existing studies typically involve linking semantic concepts to units or layers of DNNs, but fail to explain the inference process. In this paper, we introduce neural architecture disentanglement (NAD) to fill the gap. Specifically, NAD learns to disentangle a pre-trained DNN into sub-architectures according to independent tasks, forming information flows that describe the inference processes. We investigate whether, where, and how the disentanglement occurs through experiments conducted with handcrafted and automatically-searched network architectures, on both object-based and scene-based datasets. Based on the experimental results, we present three new findings that provide fresh insights into the inner logic of DNNs. First, DNNs can be divided into sub-architectures for independent tasks. Second, deeper layers do not always correspond to higher semantics. Third, the connection type in a DNN affects how the information flows across layers, leading to different disentanglement behaviors. With NAD, we further explain why DNNs sometimes give wrong predictions. Experimental results show that misclassified images have a high probability of being assigned to task sub-architectures similar to the correct ones. Our code is available at https://github.com/hujiecpp/NAD.

Existing RGB-D saliency detection models do not explicitly encourage RGB and depth to achieve effective multi-modal learning. In this paper, we introduce a novel multistage cascaded learning framework via mutual information minimization to explicitly model the multi-modal information between RGB image and depth data. Specifically, we first map the feature of each mode to a lower dimensional feature vector, and adopt mutual information minimization as a regularizer to reduce the redundancy between appearance features from RGB and geometric features from depth. We then perform multi-stage cascaded learning to impose the mutual information minimization constraint at every stage of the network. Extensive experiments on benchmark RGB-D saliency datasets illustrate the effectiveness of our framework. Further, to prosper the development of this field, we contribute the largest (7× larger than NJU2K) COME15K dataset, which contains 15,625 image pairs with high quality polygon-/scribble-/object-/instance-/rank-level annotations. Based on these rich labels, we additionally construct four new benchmarks with strong baselines and observe some interesting phenomena, which can motivate future model design. Source code and dataset are available at https://github.com/JingZhang617/cascaded_rgbd_sod.

Annotation burden has become one of the biggest barriers to semantic segmentation. Approaches based on click-level annotations have therefore attracted increasing attention due to their superior trade-off between supervision and annotation cost. In this paper, we propose seminar learning, a new learning paradigm for semantic segmentation with click-level supervision. The fundamental rationale of seminar learning is to leverage the knowledge from different networks to compensate for insufficient information provided in click-level annotations. Mimicking a seminar, our seminar learning involves a teacher-student and a student-student module, where a student can learn from both skillful teachers and other students. The teacher-student module uses a teacher network based on the exponential moving average to guide the training of the student network. In the student-student module, heterogeneous pseudo-labels are proposed to bridge the transfer of knowledge among students to enhance each other’s performance. Experimental results demonstrate the effectiveness of seminar learning, which achieves the new state-of-the-art performance of 72.51% (mIOU), surpassing previous methods by a large margin of up to 16.88% on the Pascal VOC 2012 dataset.

Appearance and motion are two important sources of information in video object segmentation (VOS). Previous methods mainly focus on using simplex solutions, lowering the upper bound of feature collaboration among and across these two cues. In this paper, we study a novel framework, termed the FSNet (Full-duplex Strategy Network), which designs a relational cross-attention module (RCAM) to achieve the bidirectional message propagation across embedding subspaces. Furthermore, the bidirectional purification module (BPM) is introduced to update the inconsistent features between the spatial-temporal embeddings, effectively improving the model robustness. By considering the mutual restraint within the full-duplex strategy, our FSNet performs the cross-modal feature-passing (i.e., transmission and receiving) simultaneously before the fusion and decoding stage, making it robust to various challenging scenarios (e.g., motion blur, occlusion) in VOS. Extensive experiments on five popular benchmarks (i.e., DAVIS16, FBMS, MCL, SegTrack-V2, and DAVSOD19) show that our FSNet outperforms other state-of-the-arts for both the VOS and video salient object detection tasks.

Existing state-of-the-art saliency detection methods heavily rely on CNN-based architectures. Alternatively, we rethink this task from a convolution-free sequence-to-sequence perspective and predict saliency by modeling long-range dependencies, which can not be achieved by convolution. Specifically, we develop a novel unified model based on a pure transformer, namely, Visual Saliency Transformer (VST), for both RGB and RGB-D salient object detection (SOD). It takes image patches as inputs and leverages the transformer to propagate global contexts among image patches. Unlike conventional architectures used in Vision Transformer (ViT), we leverage multi-level token fusion and propose a new token upsampling method under the transformer framework to get high-resolution detection results. We also develop a token-based multi-task decoder to simultaneously perform saliency and boundary detection by introducing task-related tokens and a novel patch-task-attention mechanism. Experimental results show that our model outperforms existing methods on both RGB and RGB-D SOD benchmark datasets. Most importantly, our whole framework not only provides a new perspective for the SOD field but also shows a new paradigm for transformer-based dense prediction models. Code is available at https://github.com/nnizhang/VST.

The application of light field data in salient object detection is becoming increasingly popular recently. The difficulty lies in how to effectively fuse the features within the focal stack and how to cooperate them with the feature of the all-focus image. Previous methods usually fuse focal stack features via convolution or ConvLSTM, which are both less effective and ill-posed. In this paper, we model the information fusion within focal stack via graph networks. They introduce powerful context propagation from neighbouring nodes and also avoid ill-posed implementations. On the one hand, we construct local graph connections thus avoiding prohibitive computational costs of traditional graph networks. On the other hand, instead of processing the two kinds of data separately, we build a novel dual graph model to guide the focal stack fusion process using all-focus patterns. To handle the second difficulty, previous methods usually implement one-shot fusion for focal stack and all-focus features, hence lacking a thorough exploration of their supplements. We introduce a reciprocative guidance scheme and enable mutual guidance between these two kinds of information at multiple steps. As such, both kinds of features can be enhanced iteratively, finally benefiting the saliency prediction. Extensive experimental results show that the proposed models are all beneficial and we achieve significantly better results than state-of-the-art methods.

This work proposes a weakly-supervised temporal action localization framework, called D2-Net, which strives to temporally localize actions using video-level supervision. Our main contribution is the introduction of a novel loss formulation, which jointly enhances the discriminability of latent embeddings and robustness of the output temporal class activations with respect to foreground-background noise caused by weak supervision. The proposed formulation comprises a discriminative and a denoising loss term for enhancing temporal action localization. The discriminative term incorporates a classification loss and utilizes a top-down attention mechanism to enhance the separability of latent foreground-background embeddings. The denoising loss term explicitly addresses the foreground-background noise in class activations by simultaneously maximizing intra-video and inter-video mutual information using a bottom-up attention mechanism. As a result, activations in the foreground regions are emphasized whereas those in the background regions are suppressed, thereby leading to more robust predictions. Comprehensive experiments are performed on multiple benchmarks, including THUMOS14 and ActivityNet1.2. Our D2-Net performs favorably in comparison to the existing methods on all datasets, achieving gains as high as 2.3% in terms of mAP at IoU=0.5 on THUMOS14. Source code is available at https://github.com/naraysa/D2-Net.

Multi-label zero-shot learning (ZSL) is a more realistic counter-part of standard single-label ZSL since several objects can co-exist in a natural image. However, the occurrence of multiple objects complicates the reasoning and requires region-specific processing of visual features to preserve their contextual cues. We note that the best existing multi-label ZSL method takes a shared approach towards attending to region features with a common set of attention maps for all the classes. Such shared maps lead to diffused attention, which does not discriminatively focus on relevant locations when the number of classes are large. Moreover, mapping spatially-pooled visual features to the class semantics leads to inter-class feature entanglement, thus hampering the classification. Here, we propose an alternate approach towards region-based discriminability-preserving multi-label zero-shot classification. Our approach maintains the spatial resolution to preserve region-level characteristics and utilizes a bi-level attention module (BiAM) to enrich the features by incorporating both region and scene context information. The enriched region-level features are then mapped to the class semantics and only their class predictions are spatially pooled to obtain image-level predictions, thereby keeping the multi-class features disentangled. Our approach sets a new state of the art on two large-scale multi-label zero-shot benchmarks: NUS-WIDE and Open Images. On NUS-WIDE, our approach achieves an absolute gain of 6.9% mAP for ZSL, compared to the best published results. Source code is available at https://github.com/akshitac8/BiAM.

Although convolutional neural networks (CNNs) have achieved great success in computer vision, this work investigates a simpler, convolution-free backbone network use-fid for many dense prediction tasks. Unlike the recently-proposed Vision Transformer (ViT) that was designed for image classification specifically, we introduce the Pyramid Vision Transformer (PVT), which overcomes the difficulties of porting Transformer to various dense prediction tasks. PVT has several merits compared to current state of the arts. (1) Different from ViT that typically yields low-resolution outputs and incurs high computational and memory costs, PVT not only can be trained on dense partitions of an image to achieve high output resolution, which is important for dense prediction, but also uses a progressive shrinking pyramid to reduce the computations of large feature maps. (2) PVT inherits the advantages of both CNN and Transformer, making it a unified backbone for various vision tasks without convolutions, where it can be used as a direct replacement for CNN backbones. (3) We validate PVT through extensive experiments, showing that it boosts the performance of many downstream tasks, including object detection, instance and semantic segmentation. For example, with a comparable number of parameters, PVT+RetinaNet achieves 40.4 AP on the COCO dataset, surpassing ResNet50+RetinNet (36.3 AP) by 4.1 absolute AP (see Figure 2). We hope that PVT could, serre as an alternative and useful backbone for pixel-level predictions and facilitate future research.

Few-shot semantic segmentation (FSS) is an important task for novel (unseen) object segmentation under the data-scarcity scenario. However, most FSS methods rely on unidirectional feature aggregation, e.g., from support prototypes to get the query prediction, and from high-resolution features to guide the low-resolution ones. This usually fails to fully capture the cross-resolution feature relationships and thus leads to inaccurate estimates of the query objects. To resolve the above dilemma, we propose a cyclic memory network (CMN) to directly learn to read abundant support information from all resolution features in a cyclic manner. Specifically, we first generate N pairs (key and value) of multi-resolution query features guided by the support feature and its mask. Next, we circularly take one pair of these features as the query to be segmented, and the rest N-1 pairs are written into an external memory accordingly, i.e., this leave-one-out process is conducted for N times. In each cycle, the query feature is updated by collaboratively matching its key and value with the memory, which can elegantly cover all the spatial locations from different resolutions. Furthermore, we incorporate the query feature re-adding and the query feature recursive updating mechanisms into the memory reading operation. CMN, equipped with these merits, can thus capture cross-resolution relationships and better handle the object appearance and scale variations in FSS. Experiments on PASCAL-5iand COCO-20iwell validate the effectiveness of our model for FSS.

Accurate lighting estimation is challenging yet critical to many computer vision and computer graphics tasks such as high-dynamic-range (HDR) relighting. Existing approaches model lighting in either frequency domain or spatial domain which is insufficient to represent the complex lighting conditions in scenes and tends to produce inaccurate estimation. This paper presents NeedleLight, a new lighting estimation model that represents illumination with needlets and allows lighting estimation in both frequency domain and spatial domain jointly. An optimal thresholding function is designed to achieve sparse needlets which trims redundant lighting parameters and demonstrates superior localization properties for illumination representation. In addition, a novel spherical transport loss is designed based on optimal transport theory which guides to regress lighting representation parameters with consideration of the spatial information. Furthermore, we propose a new metric that is concise yet effective by directly evaluating the estimated illumination maps rather than rendered images. Extensive experiments show that NeedleLight achieves superior lighting estimation consistently across multiple evaluation metrics as compared with state-of-the-art methods.

Humans perform co-saliency detection by first summarizing the consensus knowledge in the whole group and then searching corresponding objects in each image. Previous methods usually lack robustness, scalability, or stability for the first process and simply fuse consensus features with image features for the second process. In this paper, we propose a novel consensus-aware dynamic convolution model to explicitly and effectively perform the "summarize and search" process. To summarize consensus image features, we first summarize robust features for every single image using an effective pooling method and then aggregate cross-image consensus cues via the self-attention mechanism. By doing this, our model meets the scalability and stability requirements. Next, we generate dynamic kernels from consensus features to encode the summarized consensus knowledge. Two kinds of kernels are generated in a supplementary way to summarize fine-grained image-specific consensus object cues and the coarse group-wise common knowledge, respectively. Then, we can effectively perform object searching by employing dynamic convolution at multiple scales. Besides, a novel and effective data synthesis method is also proposed to train our network. Experimental results on four benchmark datasets verify the effectiveness of our proposed method. Our code and saliency maps are available at https://github.com/nnizhang/CADC.

We study Universal Domain Adaptation (UniDA) problem, which is recently proposed. Different from existing domain adaptation (DA) methods, e.g., Closed set, Open set and Partial DA, UniDA does not need any prior knowledge about the overlap across the source and target label sets. We have two challenges in UniDA problem: i) Domain shift. ii) Category shift. Towards tackling above challenges, we formulate a universal adaptation network called Dual Contrastive Network (DCN), where a contrastive module and a transferability rule are included. The experimental results reveal that DCN can work stably on different UniDA settings and exceeds the state-of-the-art performance across five benchmarks against existing DA methods.

This paper introduces kernel continual learning, a simple but effective variant of continual learning that leverages the non-parametric nature of kernel methods to tackle catastrophic forgetting. We deploy an episodic memory unit that stores a subset of samples for each task to learn task-specific classifiers based on kernel ridge regression. This does not require memory replay and systematically avoids task interference in the classifiers. We further introduce variational random features to learn a data-driven kernel for each task. To do so, we formulate kernel continual learning as a variational inference problem, where a random Fourier basis is incorporated as the latent variable. The variational posterior distribution over the random Fourier basis is inferred from the coreset of each task. In this way, we are able to generate more informative kernels specific to each task, and, more importantly, the coreset size can be reduced to achieve more compact memory, resulting in more efficient continual learning based on episodic memory. Extensive evaluation on four benchmarks demonstrates the effectiveness and promise of kernels for continual learning.

Domain generalization is challenging due to the domain shift and the uncertainty caused by the inaccessibility of target domain data. In this paper, we address both challenges with a probabilistic framework based on variational Bayesian inference, by incorporating uncertainty into neural network weights. We couple domain invariance in a probabilistic formula with the variational Bayesian inference. This enables us to explore domain-invariant learning in a principled way. Specifically, we derive domain-invariant representations and classifiers, which are jointly established in a two-layer Bayesian neural network. We empirically demonstrate the effectiveness of our proposal on four widely used cross-domain visual recognition benchmarks. Ablation studies validate the synergistic benefits of our Bayesian treatment when jointly learning domain-invariant representations and classifiers for domain generalization. Further, our method consistently delivers state-of-the-art mean accuracy on all benchmarks.

Semantic segmentation has attracted increasing attention due to its important role in self-driving, and it is often realized by supervised learning with large number of well labeled maps. However, the labeled images are hard to be obtained in most circumstances, and the common way for unsupervised semantic segmentation is usually implemented by transferring the knowledge from source supervised domain to target unsupervised domain. Most researches focus on encouraging target predictions to be closer to the source ones through a weight-sharing network, and achieve certain performance. However, these methods often suffer from the domain shift problem that the networks are often trained towards the source domain and lead to performance degradation. In this paper, we propose a target-targeted domain adaptation approach by focusing the training on target domain. Our model consists of two components: the Image-to-image Translation (IIT) module to translate the source image to target domain and the Target-targeted Segmentation Adaptation (TSA) module to focus the semantic segmentation on target domain. The IIT module deals with image space alignment while the TSA module bridges the domain gap at the segmentation map level. In addition, we design a closed-loop learning to promote each other by employing feedback from TSA to IIT. Extensive experiments on GTA5 and SYNTHIA to Cityscapes demonstrate the effectiveness of our method in domain adaptation of unsupervised semantic segmentation.

Most of the existing single-stage and two-stage 3D object detectors are anchor-based methods, while the efficient but challenging anchor-free single-stage 3D object detection is not well investigated. Recent studies on 2D object detection show that the anchor-free methods also are of great potential. However, the unordered and sparse properties of point clouds prevent us from directly leveraging the advanced 2D methods on 3D point clouds. We overcome this by converting the voxel-based sparse 3D feature volumes into the sparse 2D feature maps. We propose an attentive module to fit the sparse feature maps to dense mostly on the object regions through the deformable convolution tower and the supervised mask-guided attention. By directly regressing the 3D bounding box from the enhanced and dense feature maps, we construct a novel single-stage 3D detector for point clouds in an anchor-free manner. We propose an IoU-based detection confidence re-calibration scheme to improve the correlation between the detection confidence score and the accuracy of the bounding box regression. Our code is publicly available at https://github.com/jialeli1/MGAF-3DSSD.

In this paper, we present an Intersection-over-Union (IoU) guided two-stage 3D object detector with a voxel-to-point decoder. To preserve the necessary information from all raw points and maintain the high box recall in voxel based Region Proposal Network (RPN), we propose a residual voxel-to-point decoder to extract the point features in addition to the map-view features from the voxel based RPN. We use a 3D Region of Interest (RoI) alignment to crop and align the features with the proposal boxes for accurately perceiving the object position. The RoI-Aligned features are finally aggregated with the corner geometry embeddings that can provide the potentially missing corner information in the box refinement stage. We propose a simple and efficient method to align the estimated IoUs to the refined proposal boxes as a more relevant localization confidence. The comprehensive experiments on KITTI and Waymo Open Dataset demonstrate that our method achieves significant improvements with novel architectures against the existing methods. The code is available on Github URLhttps://github.com/jialeli1/From-Voxel-to-Point .

Zero-shot learning (ZSL) tackles the unseen class recognition problem, transferring semantic knowledge from seen classes to unseen ones. Typically, to guarantee desirable knowledge transfer, a common (latent) space is adopted for associating the visual and semantic domains in ZSL. However, existing common space learning methods align the semantic and visual domains by merely mitigating distribution disagreement through one-step adaptation. This strategy is usually ineffective due to the heterogeneous nature of the feature representations in the two domains, which intrinsically contain both distribution and structure variations. To address this and advance ZSL, we propose a novel hierarchical semantic-visual adaptation (HSVA) framework. Specifically, HSVA aligns the semantic and visual domains by adopting a hierarchical two-step adaptation, i.e., structure adaptation and distribution adaptation. In the structure adaptation step, we take two task-specific encoders to encode the source data (visual domain) and the target data (semantic domain) into a structure-aligned common space. To this end, a supervised adversarial discrepancy (SAD) module is proposed to adversarially minimize the discrepancy between the predictions of two task-specific classifiers, thus making the visual and semantic feature manifolds more closely aligned. In the distribution adaptation step, we directly minimize the Wasserstein distance between the latent multivariate Gaussian distributions to align the visual and semantic distributions using a common encoder. Finally, the structure and distribution adaptation are derived in a unified framework under two partially-aligned variational autoencoders. Extensive experiments on four benchmark datasets demonstrate that HSVA achieves superior performance on both conventional and generalized ZSL. The code is available at \url{https://github.com/shiming-chen/HSVA}.

Transformers have recently gained increasing attention in computer vision. However, existing studies mostly use Transformers for feature representation learning, e.g. for image classification and dense predictions, and the generalizability of Transformers is unknown. In this work, we further investigate the possibility of applying Transformers for image matching and metric learning given pairs of images. We find that the Vision Transformer (ViT) and the vanilla Transformer with decoders are not adequate for image matching due to their lack of image-to-image attention. Thus, we further design two naive solutions, i.e. query-gallery concatenation in ViT, and query-gallery cross-attention in the vanilla Transformer. The latter improves the performance, but it is still limited. This implies that the attention mechanism in Transformers is primarily designed for global feature aggregation, which is not naturally suitable for image matching. Accordingly, we propose a new simplified decoder, which drops the full attention implementation with the softmax weighting, keeping only the query-key similarity computation. Additionally, global max pooling and a multilayer perceptron (MLP) head are applied to decode the matching result. This way, the simplified decoder is computationally more efficient, while at the same time more effective for image matching. The proposed method, called TransMatcher, achieves state-of-the-art performance in generalizable person re-identification, with up to 6.1% and 5.7% performance gains in Rank-1 and mAP, respectively, on several popular datasets. Code is available at https://github.com/ShengcaiLiao/QAConv.

Recent advances in self-supervised learning with instance-level contrastive objectives facilitate unsupervised clustering. However, a standalone datum is not perceiving the context of the holistic cluster, and may undergo sub-optimal assignment. In this paper, we extend the mainstream contrastive learning paradigm to a cluster-level scheme, where all the data subjected to the same cluster contribute to a unified representation that encodes the context of each data group. Contrastive learning with this representation then rewards the assignment of each datum. To implement this vision, we propose twin-contrast clustering (TCC). We define a set of categorical variables as clustering assignment confidence, which links the instance-level learning track with the cluster-level one. On one hand, with the corresponding assignment variables being the weight, a weighted aggregation along the data points implements the set representation of a cluster. We further propose heuristic cluster augmentation equivalents to enable cluster-level contrastive learning. On the other hand, we derive the evidence lower-bound of the instance-level contrastive objective with the assignments. By reparametrizing the assignment variables, TCC is trained end-to-end, requiring no alternating steps. Extensive experiments show that TCC outperforms the state-of-the-art on benchmarked datasets.

Multi-task learning aims to explore task relatedness to improve individual tasks, which is of particular significance in the challenging scenario that only limited data is available for each task. To tackle this challenge, we propose variational multi-task learning (VMTL), a general probabilistic inference framework for learning multiple related tasks. We cast multi-task learning as a variational Bayesian inference problem, in which task relatedness is explored in a unified manner by specifying priors. To incorporate shared knowledge into each task, we design the prior of a task to be a learnable mixture of the variational posteriors of other related tasks, which is learned by the Gumbel-Softmax technique. In contrast to previous methods, our VMTL can exploit task relatedness for both representations and classifiers in a principled way by jointly inferring their posteriors. This enables individual tasks to fully leverage inductive biases provided by related tasks, therefore improving the overall performance of all tasks. Experimental results demonstrate that the proposed VMTL is able to effectively tackle a variety of challenging multi-task learning settings with limited training data for both classification and regression. Our method consistently surpasses previous methods, including strong Bayesian approaches, and achieves state-of-the-art performance on five benchmark datasets.

In this paper, we propose variational prototype inference to address few-shot semantic segmentation in a probabilistic framework. A probabilistic latent variable model infers the distribution of the prototype that is treated as the latent variable. We formulate the optimization as a variational inference problem, which is established with an amortized inference network based on an auto-encoder architecture. The probabilistic modeling of the prototype enhances its generalization ability to handle the inherent uncertainty caused by limited data and the huge intra-class variations of objects. Moreover, it offers a principled way to incorporate the prototype extracted from support images into the prediction of the segmentation maps for query images. We conduct extensive experimental evaluations on three benchmark datasets. Ablation studies show the effectiveness of variational prototype inference for few-shot semantic segmentation by probabilistic modeling. On all three benchmarks, our proposal achieves high segmentation accuracy and surpasses previous methods by considerable margins.

Three-dimensional (3D) human pose estimation involves estimating the articulated 3D joint locations of a human body from an image or video. Due to its widespread applications in a great variety of areas, such as human motion analysis, human–computer interaction, robots, 3D human pose estimation has recently attracted increasing attention in the computer vision community, however, it is a challenging task due to depth ambiguities and the lack of in-the-wild datasets. A large number of approaches, with many based on deep learning, have been developed over the past decade, largely advancing the performance on existing benchmarks. To guide future development, a comprehensive literature review is highly desired in this area. However, existing surveys on 3D human pose estimation mainly focus on traditional methods and a comprehensive review on deep learning based methods remains lacking in the literature. In this paper, we provide a thorough review of existing deep learning based works for 3D pose estimation, summarize the advantages and disadvantages of these methods and provide an in-depth understanding of this area. Furthermore, we also explore the commonly-used benchmark datasets on which we conduct a comprehensive study for comparison and analysis. Our study sheds light on the state of research development in 3D human pose estimation and provides insights that can facilitate the future design of models and algorithms.

Salient object detection, which simulates human visual perception in locating the most significant object(s) in a scene, has been widely applied to various computer vision tasks. Now, the advent of depth sensors means that depth maps can easily be captured; this additional spatial information can boost the performance of salient object detection. Although various RGB-D based salient object detection models with promising performance have been proposed over the past several years, an in-depth understanding of these models and the challenges in this field remains lacking. In this paper, we provide a comprehensive survey of RGB-D based salient object detection models from various perspectives, and review related benchmark datasets in detail. Further, as light fields can also provide depth maps, we review salient object detection models and popular benchmark datasets from this domain too. Moreover, to investigate the ability of existing models to detect salient objects, we have carried out a comprehensive attribute-based evaluation of several representative RGB-D based salient object detection models. Finally, we discuss several challenges and open directions of RGB-D based salient object detection for future research. All collected models, benchmark datasets, datasets constructed for attribute-based evaluation, and related code are publicly available at https://github.com/taozh2017/RGBD-SODsurvey.

Hyperparameters are numerical pre-sets whose values are assigned prior to the commencement of a learning process. Selecting appropriate hyperparameters is often critical for achieving satisfactory performance in many vision problems, such as deep learning-based visual object tracking. However, it is often difficult to determine their optimal values, especially if they are specific to each video input. Most hyperparameter optimization algorithms tend to search a generic range and are imposed blindly on all sequences. In this paper, we propose a novel dynamical hyperparameter optimization method that adaptively optimizes hyperparameters for a given sequence using an action-prediction network leveraged on continuous deep Q-learning. Since the observation space for object tracking is significantly more complex than those in traditional control problems, existing continuous deep Q-learning algorithms cannot be directly applied. To overcome this challenge, we introduce an efficient heuristic strategy to handle high dimensional state space, while also accelerating the convergence behavior. The proposed algorithm is applied to improve two representative trackers, a Siamese-based one and a correlation-filter-based one, to evaluate its generalizability. Their superior performances on several popular benchmarks are clearly demonstrated. Our source code is available at https://github.com/shenjianbing/dqltracking.

Deep neural networks can easily be fooled by an adversary with minuscule perturbations added to an input image. The existing defense techniques suffer greatly under white-box attack settings, where an adversary has full knowledge of the network and can iterate several times to find strong perturbations. We observe that the main reason for the existence of such vulnerabilities is the close proximity of different class samples in the learned feature space of deep models. This allows the model decisions to be completely changed by adding an imperceptible perturbation to the inputs. To counter this, we propose to class-wise disentangle the intermediate feature representations of deep networks, specifically forcing the features for each class to lie inside a convex polytope that is maximally separated from the polytopes of other classes. In this manner, the network is forced to learn distinct and distant decision regions for each class. We observe that this simple constraint on the features greatly enhances the robustness of learned models, even against the strongest white-box attacks, without degrading the classification performance on clean images. We report extensive evaluations in both black-box and white-box attack scenarios and show significant gains in comparison to state-of-the-art defenses.

The self-expressive property of data points, that is, each data point can be linearly represented by the other data points in the same subspace, has proven effective in leading subspace clustering (SC) methods. Most self-expressive methods usually construct a feasible affinity matrix from a coefficient matrix, obtained by solving an optimization problem. However, the negative entries in the coefficient matrix are forced to be positive when constructing the affinity matrix via exponentiation, absolute symmetrization, or squaring operations. This consequently damages the inherent correlations among the data. Besides, the affine constraint used in these methods is not flexible enough for practical applications. To overcome these problems, in this article, we introduce a scaled simplex representation (SSR) for the SC problem. Specifically, the non-negative constraint is used to make the coefficient matrix physically meaningful, and the coefficient vector is constrained to be summed up to a scalar to make it more discriminative. The proposed SSR-based SC (SSRSC) model is reformulated as a linear equality-constrained problem, which is solved efficiently under the alternating direction method of multipliers framework. Experiments on benchmark datasets demonstrate that the proposed SSRSC algorithm is very efficient and outperforms the state-of-the-art SC methods on accuracy. The code can be found at https://github.com/csjunxu/SSRSC.

Accurately recognizing different categories of sceneries with sophisticated spatial configurations is a useful technique in computer vision and intelligent systems, e.g., scene understanding and autonomous driving. Competitive accuracies have been observed by the deep recognition models recently. Nevertheless, these deep architectures cannot explicitly characterize human visual perception, that is, the sequence of gaze allocation and the subsequent cognitive processes when viewing each scenery. In this paper, a novel spatially aware aggregation network is proposed for scene categorization, where the human gaze behavior is discovered in a semisupervised setting. In particular, as semantically labeling a large quantity of scene images is labor-intensive, a semisupervised and structure-preserved non-negative matrix factorization (NMF) is proposed to detect a set of visually/semantically salient regions from each scenery. Afterward, the gaze shifting path (GSP) is engineered to characterize the process of humans perceiving each scene picture. To deeply describe each GSP, a novel spatially aware CNN termed SA-Net is developed. It accepts input regions with various shapes and statistically aggregates all the salient regions along each GSP. Finally, the learned deep GSP features from the entire scene images are fused into an image kernel, which is subsequently integrated into a kernel SVM to categorize different sceneries. Comparative experiments on six scene image sets have shown the advantage of our method.

Learning discriminative and invariant feature representation is the key to visual image categorization. In this article, we propose a novel invariant deep compressible covariance pooling (IDCCP) to solve nuisance variations in aerial scene categorization. We consider transforming the input image according to a finite transformation group that consists of multiple confounding orthogonal matrices, such as the D4 group. Then, we adopt a Siamese-style network to transfer the group structure to the representation space, where we can derive a trivial representation that is invariant under the group action. The linear classifier trained with trivial representation will also be possessed with invariance. To further improve the discriminative power of representation, we extend the representation to the tensor space while imposing orthogonal constraints on the transformation matrix to effectively reduce feature dimensions. We conduct extensive experiments on the publicly released aerial scene image data sets and demonstrate the superiority of this method compared with state-of-the-art methods. In particular, with using ResNet architecture, our IDCCP model can reduce the dimension of the tensor representation by about 98% without sacrificing accuracy (i.e., < 0.5%).

Matching person images between the daytime visible modality and night-time infrared modality (VI-ReID) is a challenging cross-modality pedestrian retrieval problem. Existing methods usually learn the multi-modality features in raw image, ignoring the image-level discrepancy. Some methods apply GAN technique to generate the cross-modality images, but it destroys the local structure and introduces unavoidable noise. In this paper, we propose a Homogeneous Augmented Tri-Modal (HAT) learning method for VI-ReID, where an auxiliary grayscale modality is generated from their homogeneous visible images, without additional training process. It preserves the structure information of visible images and approximates the image style of infrared modality. Learning with the grayscale visible images enforces the network to mine structure relations across multiple modalities, making it robust to color variations. Specifically, we solve the tri-modal feature learning from both multi-modal classification and multi-view retrieval perspectives. For multi-modal classification, we learn a multi-modality sharing identity classifier with a parameter-sharing network, trained with a homogeneous and heterogeneous identification loss. For multi-view retrieval, we develop a weighted tri-directional ranking loss to optimize the relative distance across multiple modalities. Incorporated with two invariant regularizers, HAT simultaneously minimizes multiple modality variations. In-depth analysis demonstrates the homogeneous grayscale augmentation significantly outperforms the current state-of-the-art by a large margin.

Deep learning has recently been intensively studied in the context of image compressive sensing (CS) to discover and represent complicated image structures. These approaches, however, either suffer from nonflexibility for an arbitrary sampling ratio or lack an explicit deep-learned regularization term. This paper aims to solve the CS reconstruction problem by combining the deep-learned regularization term and proximal operator. We first introduce a regularization term using a carefully designed residual-regressive net, which can measure the distance between a corrupted image and a clean image set and accurately identify to which subspace the corrupted image belongs. We then address a proximal operator with a tailored dilated residual channel attention net, which enables the learned proximal operator to map the distorted image into the clean image set. We adopt an adaptive proximal selection strategy to embed the network into the loop of the CS image reconstruction algorithm. Moreover, a self-ensemble strategy is presented to improve CS recovery performance. We further utilize state evolution to analyze the effectiveness of the designed networks. Extensive experiments also demonstrate that our method can yield superior accurate reconstruction (PSNR gain over 1 dB) compared to other competing approaches while achieving the current state-of-the-art image CS reconstruction performance. The test code is available at https://github.com/zjut-gwl/CSDRCANet.

Due to the advantages of real-time detection and improved performance, single-shot detectors have gained great attention recently. To solve the complex scale variations, single-shot detectors make scale-aware predictions based on multiple pyramid layers. Typically, small objects are detected on shallow layers while large objects are detected on deep layers. However, the features in the pyramid are not scale-aware enough, which limits the detection performance. Two common problems in single-shot detectors caused by object scale variations can be observed: (1) false negative problem, i.e., small objects are easily missed due to the weak features; (2) part-false positive problem, i.e., the salient part of a large object is sometimes detected as an object. With this observation, a new Neighbor Erasing and Transferring (NET) mechanism is proposed for feature scale-unmixing to explore scale-aware features in this paper. In NET, a Neighbor Erasing Module (NEM) is designed to erase the salient features of large objects and emphasize the features of small objects in shallow layers. A Neighbor Transferring Module (NTM) is introduced to transfer the erased features and highlight large objects in deep layers. With this mechanism, a single-shot network called NETNet is constructed for scale-aware object detection. In addition, we propose to aggregate nearest neighboring pyramid features to enhance our NET. Experiments on MS COCO dataset and UAVDT dataset demonstrate the effectiveness of our method. NETNet obtains 38.5% AP at a speed of 27 FPS and 32.0% AP at a speed of 55 FPS on MS COCO dataset. As a result, NETNet achieves a better trade-off for real-time and accurate object detection.

Different from conventional instance segmentation, salient instance segmentation (SIS) faces two difficulties. The first is that it involves segmenting salient instances only while ignoring background, and the second is that it targets generic object instances without pre-defined object categories. In this paper, based on the state-of-the-art Mask R-CNN model, we propose to leverage complementary saliency and contour information to handle these two challenges. We first improve Mask R-CNN by introducing an interleaved execution strategy and proposing a novel mask head network to incorporate global context within each RoI. Then we add two branches to Mask R-CNN for saliency and contour detection, respectively. We fuse the Mask R-CNN features with the saliency and contour features, where the former supply pixel-wise saliency information to help with identifying salient regions and the latter provide a generic object contour prior to help detect and segment generic objects. We also propose a novel multiscale global attention model to generate attentive global features from multiscale representative features for feature fusion. Experimental results demonstrate that all our proposed model components can improve SIS performance. Finally, our overall model outperforms state-of-the-art SIS methods and Mask R-CNN by more than 6% and 3%, respectively. By using additional multitask training data, we can further improve the model performance on the ILSO dataset.

Deep convolutional neural networks have largely benefited computer vision tasks. However, the high computational complexity limits their real-world applications. To this end, many methods have been proposed for efficient network learning, and applications in portable mobile devices. In this paper, we propose a novel Moving-Mobile-Network, named M2Net, for landmark recognition, equipped each landmark image with located geographic information. We intuitively find that M2Net can essentially promote the diversity of the inference path (selected blocks subset) selection, so as to enhance the recognition accuracy. The above intuition is achieved by our proposed reward function with the input of geo-location and landmarks. We also find that the performance of other portable networks can be improved via our architecture. We construct two landmark image datasets, with each landmark associated with geographic information, over which we conduct extensive experiments to demonstrate that M2Net achieves improved recognition accuracy with comparable complexity.

Unsupervised domain adaptation for object detection is a challenging problem with many real-world applications. Unfortunately, it has received much less attention than supervised object detection. Models that try to address this task tend to suffer from a shortage of annotated training samples. Moreover, existing methods of feature alignments are not sufficient to learn domain-invariant representations. To address these limitations, we propose a novel augmented feature alignment network (AFAN) which integrates intermediate domain image generation and domain-adversarial training into a unified framework. An intermediate domain image generator is proposed to enhance feature alignments by domain-adversarial training with automatically generated soft domain labels. The synthetic intermediate domain images progressively bridge the domain divergence and augment the annotated source domain training data. A feature pyramid alignment is designed and the corresponding feature discriminator is used to align multi-scale convolutional features of different semantic levels. Last but not least, we introduce a region feature alignment and an instance discriminator to learn domain-invariant features for object proposals. Our approach significantly outperforms the state-of-the-art methods on standard benchmarks for both similar and dissimilar domain adaptations. Further extensive experiments verify the effectiveness of each component and demonstrate that the proposed network can learn domain-invariant representations.

Traditional operations, e.g. graph edit distance (GED), are no longer suitable for processing the massive quantities of graph-structured data now available, due to their irregular structures and high computational complexities. With the advent of graph neural networks (GNNs), the problems of graph representation and graph similarity search have drawn particular attention in the field of computer vision. However, GNNs have been less studied for efficient and fast retrieval after graph representation. To represent graph-based data, and maintain fast retrieval while doing so, we introduce an efficient hash model with graph neural networks (HGNN) for a newly designed task (i.e. fast graph-based data retrieval). Due to its flexibility, HGNN can be implemented in both an unsupervised and supervised manner. Specifically, by adopting a graph neural network and hash learning algorithms, HGNN can effectively learn a similarity-preserving graph representation and compute pair-wise similarity or provide classification via low-dimensional compact hash codes. To the best of our knowledge, our model is the first to address graph hashing representation in the Hamming space. Our experimental results reach comparable prediction accuracy to full-precision methods and can even outperform traditional models in some cases. In real-world applications, using hash codes can greatly benefit systems with smaller memory capacities and accelerate the retrieval speed of graph-structured data. Hence, we believe the proposed HGNN has great potential in further research.

Pedestrian detection relying on deep convolution neural networks has made significant progress. Though promising results have been achieved on standard pedestrians, the performance on heavily occluded pedestrians remains far from satisfactory. The main culprits are intra-class occlusions involving other pedestrians and inter-class occlusions caused by other objects, such as cars and bicycles. These result in a multitude of occlusion patterns. We propose an approach for occluded pedestrian detection with the following contributions. First, we introduce a novel mask-guided attention network that fits naturally into popular pedestrian detection pipelines. Our attention network emphasizes on visible pedestrian regions while suppressing the occluded ones by modulating full body features. Second, we propose the occlusion-sensitive hard example mining method and occlusion-sensitive loss that mines hard samples according to the occlusion level and assigns higher weights to the detection errors occurring at highly occluded pedestrians. Third, we empirically demonstrate that weak box-based segmentation annotations provide reasonable approximation to their dense pixel-wise counterparts. Experiments are performed on CityPersons, Caltech and ETH datasets. Our approach sets a new state-of-the-art on all three datasets. Our approach obtains an absolute gain of 10.3% in log-average miss rate, compared with the best reported results on the heavily occluded HO pedestrian set of the CityPersons test set. Code and models are available at: https://github.com/Leotju/MGAN.

Transductive zero-shot learning (TZSL) extends conventional ZSL by leveraging (unlabeled) unseen images for model training. A typical method for ZSL involves learning embedding weights from the feature space to the semantic space. However, the learned weights in most existing methods are dominated by seen images, and can thus not be adapted to unseen images very well. In this paper, to align the (embedding) weights for better knowledge transfer between seen/unseen classes, we propose the virtual mainstay alignment network (VMAN), which is tailored for the transductive ZSL task. Specifically, VMAN is casted as a tied encoder-decoder net, thus only one linear mapping weights need to be learned. To explicitly learn the weights in VMAN, for the first time in ZSL, we propose to generate virtual mainstay (VM) samples for each seen class, which serve as new training data and can prevent the weights from being shifted to seen images, to some extent. Moreover, a weighted reconstruction scheme is proposed and incorporated into the model training phase, in both the semantic/feature spaces. In this way, the manifold relationships of the VM samples are well preserved. To further align the weights to adapt to more unseen images, a novel instance-category matching regularization is proposed for model re-training. VMAN is thus modeled as a nested minimization problem and is solved by a Taylor approximate optimization paradigm. In comprehensive evaluations on four benchmark datasets, VMAN achieves superior performances under the (Generalized) TZSL setting.

Multi-level feature fusion is a fundamental topic in computer vision. It has been exploited to detect, segment and classify objects at various scales. When multi-level features meet multi-modal cues, the optimal feature aggregation and multi-modal learning strategy become a hot potato. In this paper, we leverage the inherent multi-modal and multi-level nature of RGB-D salient object detection to devise a novel Bifurcated Backbone Strategy Network (BBS-Net). Our architecture, is simple, efficient, and backbone-independent. In particular, first, we propose to regroup the multi-level features into teacher and student features using a bifurcated backbone strategy (BBS). Second, we introduce a depth-enhanced module (DEM) to excavate informative depth cues from the channel and spatial views. Then, RGB and depth modalities are fused in a complementary way. Extensive experiments show that BBS-Net significantly outperforms 18 state-of-the-art (SOTA) models on eight challenging datasets under five evaluation measures, demonstrating the superiority of our approach (~4% improvement in S-measure vs . the top-ranked model: DMRA). In addition, we provide a comprehensive analysis on the generalization ability of different RGB-D datasets and provide a powerful training set for future research. The complete algorithm, benchmark results, and post-processing toolbox are publicly available at https://github.com/zyjwuyan/BBS-Net.

Intelligently understanding the sophisticated topological structures from aerial photographs is a useful technique in aerial image analysis. Conventional methods cannot fulfill this task due to the following challenges: 1) the topology number of an aerial photo increases exponentially with the topology size, which requires a fine-grained visual descriptor to discriminatively represent each topology; 2) identifying visually/semantically salient topologies within each aerial photo in a weakly-labeled context, owing to the unaffordable human resources required for pixel-level annotation; and 3) designing a cross-domain knowledge transferal module to augment aerial photo perception, since multi-resolution aerial photos are taken asynchronistically in practice. To handle the above problems, we propose a unified framework to understand aerial photo topologies, focusing on representing each aerial photo by a set of visually/semantically salient topologies based on human visual perception and further employing them for visual categorization. Specifically, we first extract multiple atomic regions from each aerial photo, and thereby graphlets are built to capture the each aerial photo topologically. Then, a weakly-supervised ranking algorithm selects a few semantically salient graphlets by seamlessly encoding multiple image-level attributes. Toward a visualizable and perception-aware framework, we construct gaze shifting path (GSP) by linking the top-ranking graphlets. Finally, we derive the deep GSP representation, and formulate a semi-supervised and cross-domain SVM to partition each aerial photo into multiple categories. The SVM utilizes the global composition from low-resolution counterparts to enhance the deep GSP features from high-resolution aerial photos which are partially-annotated. Extensive visualization results and categorization performance comparisons have demonstrated the competitiveness of our approach.

Recent representative scene parsing methods based on Convolutional Neural Networks (CNN) have greatly improved spatial resolution of pixel-wise labelling by exploiting multi-scale features and refined boundaries. However, the vast majority of previous works only utilize the color or textural information of images, without considering the depth information, which is beneficial for semantic reasoning. In this paper, we take advantages of the mutual benefit and strong correlation between depth information and semantic information in scene parsing by introducing the Built-in Depth-Semantic Coupled Encoding (BDSCE) module, which adaptively fuses RGB and depth features, and selectively highlights the depth-discriminative features. The proposed BDSCE module is compatible with existing CNN based methods, and can greatly improve scene parsing performance, particularly in those categories that have clear depth distinction and might be misclassified with RGB-only features. Furthermore, we also extend our proposed module to other urban scene semantic reasoning tasks such as vehicle detection and road segmentation, which are implemented by effectively learning and exploiting the encoded depth-semantics and transferring the learned representations with fine-tuning. The extensive experiments on the popular datasets Cityscapes and KITTI demonstrate that our method performs quite well and can significantly improve the state-of-the-art methods.

In supervised learning for medical image analysis, sample selection methodologies are fundamental to attain optimum system performance promptly and with minimal expert interactions (e.g. label querying in an active learning setup). In this article we propose a novel sample selection methodology based on deep features leveraging information contained in interpretability saliency maps. In the absence of ground truth labels for informative samples, we use a novel self supervised learning based approach for training a classifier that learns to identify the most informative sample in a given batch of images. We demonstrate the benefits of the proposed approach, termed Interpretability-Driven Sample Selection (IDEAL), in an active learning setup aimed at lung disease classification and histopathology image segmentation. We analyze three different approaches to determine sample informativeness from interpretability saliency maps: (i) an observational model stemming from findings on previous uncertainty-based sample selection approaches, (ii) a radiomics-based model, and (iii) a novel data-driven self-supervised approach. We compare IDEAL to other baselines using the publicly available NIH chest X-ray dataset for lung disease classification, and a public histopathology segmentation dataset (GLaS), demonstrating the potential of using interpretability information for sample selection in active learning systems. Results show our proposed self supervised approach outperforms other approaches in selecting informative samples leading to state of the art performance with fewer samples.

Retinal fundus images are widely used for the clinical screening and diagnosis of eye diseases. However, fundus images captured by operators with various levels of experience have a large variation in quality. Low-quality fundus images increase uncertainty in clinical observation and lead to the risk of misdiagnosis. However, due to the special optical beam of fundus imaging and structure of the retina, natural image enhancement methods cannot be utilized directly to address this. In this article, we first analyze the ophthalmoscope imaging system and simulate a reliable degradation of major inferior-quality factors, including uneven illumination, image blurring, and artifacts. Then, based on the degradation model, a clinically oriented fundus enhancement network (cofe-Net) is proposed to suppress global degradation factors, while simultaneously preserving anatomical retinal structures and pathological characteristics for clinical observation and analysis. Experiments on both synthetic and real images demonstrate that our algorithm effectively corrects low-quality fundus images without losing retinal details. Moreover, we also show that the fundus correction method can benefit medical image analysis applications, e.g., retinal vessel segmentation and optic disc/cup detection.

People with diabetes are at risk of developing an eye disease called diabetic retinopathy (DR). This disease occurs when high blood glucose levels cause damage to blood vessels in the retina. Computer-aided DR diagnosis has become a promising tool for the early detection and severity grading of DR, due to the great success of deep learning. However, most current DR diagnosis systems do not achieve satisfactory performance or interpretability for ophthalmologists, due to the lack of training data with consistent and fine-grained annotations. To address this problem, we construct a large fine-grained annotated DR dataset containing 2,842 images (FGADR). Specifically, this dataset has 1,842 images with pixel-level DR-related lesion annotations, and 1,000 images with image-level labels graded by six board-certified ophthalmologists with intra-rater consistency. The proposed dataset will enable extensive studies on DR diagnosis. Further, we establish three benchmark tasks for evaluation: 1. DR lesion segmentation; 2. DR grading by joint classification and segmentation; 3. Transfer learning for ocular multi-disease identification. Moreover, a novel inductive transfer learning method is introduced for the third task. Extensive experiments using different state-of-the-art methods are conducted on our FGADR dataset, which can serve as baselines for future research. Our dataset will be released in https://csyizhou.github.io/FGADR/.

Automatic thoracic disease diagnosis is a rising research topic in the medical imaging community, with many potential applications. However, the inconsistent appearances and high complexities of various lesions in chest X-rays currently hinder the development of a reliable and robust intelligent diagnosis system. Attending to the high-probability abnormal regions and exploiting the priori of a related knowledge graph offers one promising route to addressing these issues. As such, in this paper, we propose two contrastive abnormal attention models and a dual-weighting graph convolution to improve the performance of thoracic multi-disease recognition. First, a left-right lung contrastive network is designed to learn intra-attentive abnormal features to better identify the most common thoracic diseases, whose lesions rarely appear in both sides symmetrically. Moreover, an inter-contrastive abnormal attention model aims to compare the query scan with multiple anchor scans without lesions to compute the abnormal attention map. Once the intra- and inter-contrastive attentions are weighted over the features, in addition to the basic visual spatial convolution, a chest radiology graph is constructed for dual-weighting graph reasoning. Extensive experiments on the public NIH ChestX-ray and CheXpert datasets show that our model achieves consistent improvements over the state-of-the-art methods both on thoracic disease identification and localization.

Recently, imagesmoothing has gained increasing attention due to its prerequisite role in other image processing tasks, e.g., image enhancement and editing. However, the evaluation of image smoothing algorithms is usually performed by subjective observation on images without corresponding ground truths. To promote the development of image smoothing algorithms, in this paper, we construct a novel Nankai Smoothing (NKS) dataset containing 200 images blended by versatile structure images and natural textures. The structure images are inherently smooth and naturally taken as ground truths. On our NKS dataset, we comprehensively evaluate 14 popular image smoothing algorithms. Moreover, we propose a Pixel-level Non-Local Smoothing (PNLS) method to well preserve the structure of the smoothed images, by exploiting the pixel-level non-local self-similarity prior of natural images. Extensive experiments on several benchmark datasets demonstrate that our PNLS outperforms previous algorithms on the image smoothing task. Ablation studies also reveal the work mechanism of our PNLS on image smoothing. To further show its effectiveness, we apply our PNLS on several applications such as semantic region smoothing, detail/edge enhancement, and image abstraction. The dataset and code are available athttps://github.com/zal0302/PNLS.

Facial landmark detection in the wild remains a challenging problem in computer vision. Deep learning-based methods currently play a leading role in solving this. However, these approaches generally focus on local feature learning and ignore global relationships. Therefore, in this study, a self-attention mechanism is introduced into facial landmark detection. Specifically, a coarse-to-fine facial landmark detection method is proposed that uses two stacked hourglasses as the backbone, with a new landmark-guided self-attention (LGSA) block inserted between them. The LGSA block learns the global relationships between different positions on the feature map and allows feature learning to focus on the locations of landmarks with the help of a landmark-specific attention map, which is generated in the first-stage hourglass model. A novel attentional consistency loss is also proposed to ensure the generation of an accurate landmark-specific attention map. A new channel transformation block is used as the building block of the hourglass model to improve the model's capacity. The coarse-to-fine strategy is adopted during and between phases to reduce complexity. Extensive experimental results on public datasets demonstrate the superiority of our proposed method against state-of-the-art models.

Human visual system can selectively attend to parts of a scene for quick perception, a biological mechanism known asHuman attention. Inspired by this, recent deep learning models encode attention mechanisms to focus on the most task-relevant parts of the input signal for further processing, which is calledMachine/Neural/Artificial attention. Understanding the relation between human and machine attention is important for interpreting and designing neural networks. Many works claim that the attention mechanism offers an extra dimension of interpretability by explaining where the neural networks look. However, recent studies demonstrate that artificial attention maps do not always coincide with common intuition. In view of these conflicting evidence, here we make a systematic study on using artificial attention and human attention in neural network design. With three example computer vision tasks (i.e., salient object segmentation, video action recognition, and fine-grained image classification), diverse representative backbones (i.e., AlexNet, VGGNet, ResNet) and famous architectures (i.e., Two-stream, FCN), corresponding real human gaze data, and systematically conducted large-scale quantitative studies, we quantify the consistency between artificial attention and human visual attention and offer novel insights into existing artificial attention mechanisms by giving preliminary answers to several key questions related to human and artificial attention mechanisms. Overall results demonstrate that human attention can benchmark the meaningful ‘ground-truth’ in attention-driven tasks, where the more the artificial attention is close to human attention, the better the performance; for higher-level vision tasks, it is case-by-case. It would be advisable for attention-driven tasks to explicitly force a better alignment between artificial and human attention to boost the performance; such alignment would also improve the network explainability for higher-level computer vision tasks.

Hashing methods have proven to be effective in the field of large-scale image retrieval. In recent years, the performance of hashing algorithms based on deep learning has greatly exceeded that of non-deep methods. However, most of the outstanding hashing methods are supervised models that heavily rely on annotated labels. In order to circumvent the huge overhead of labeling large-scale datasets, some unsupervised hashing algorithms have been proposed, such as pseudo labels and pseudo pairs. Since the image labels are strictly unavailable, some hyper-parameters in these methods are difficult to be selected, e.g., the final result is very sensitive to the picked number of categories or the chosen threshold of similarity for pairs. In addition, the calculation of pseudo-labels in high-dimensional space is not only computationally complex, but also has low precision. Therefore, in order to alleviate these issues in this paper, we propose a simple but effective Deep Unsupervised Self-evolutionary Hashing (DUSH) algorithm, which utilizes a curriculum learning strategy to iteratively select pseudo pairs from easy to hard in low dimensional Hamming space. Extensive experiments are conducted on four popular datasets, including two single-label datasets and two multi-label datasets, and the results show that our method can significantly outperform the state-of-the-art methods.

Spectral super-resolution (SSR) aims at generating a hyperspectral image (HSI) from a given RGB image. Recently, a promising direction is to learn a complicated mapping function from the RGB image to the HSI counterpart using a deep convolutional neural network. This essentially involves mapping the RGB context within a size-specific receptive field centered at each pixel to its spectrum in the HSI. The focus thereon is to appropriately determine the receptive field size and establish the mapping function from RGB context to the corresponding spectrum. Due to their differences in category or spatial position, pixels in HSIs often require different-sized receptive fields and distinct mapping functions. However, few efforts have been invested to explicitly exploit this prior.To address this problem, we propose a pixel-aware deep function-mixture network for SSR, which is composed of a new class of modules, termed function-mixture (FM) blocks. Each FM block is equipped with some basis functions, i.e., parallel subnets of different-sized receptive fields. Besides, it incorporates an extra subnet as a mixing function to generate pixel-wise weights, and then linearly mixes the outputs of all basis functions with those generated weights. This enables us to pixel-wisely determine the receptive field size and the mapping function. Moreover, we stack several such FM blocks to further increase the flexibility of the network in learning the pixel-wise mapping. To encourage feature reuse, intermediate features generated by the FM blocks are fused in late stage, which proves to be effective for boosting the SSR performance. Experimental results on three benchmark HSI datasets demonstrate the superiority of the proposed method.

The main requisite for fine-grained recognition task is to focus on subtle discriminative details that make the subordinate classes different from each other. We note that existing methods implicitly address this requirement and leave it to a data-driven pipeline to figure out what makes a subordinate class different from the others. This results in two major limitations: First, the network focuses on the most obvious distinctions between classes and overlooks more subtle inter-class variations. Second, the chance of misclassifying a given sample in any of the negative classes is considered equal, while in fact, confusions generally occur among only the most similar classes. Here, we propose to explicitly force the network to find the subtle differences among closely related classes. In this pursuit, we introduce two key novelties that can be easily plugged into existing end-to-end deep learning pipelines. On one hand, we introduce “diversification block” which masks the most salient features for an input to force the network to use more subtle cues for its correct classification. Concurrently, we introduce a “gradient-boosting” loss function that focuses only on the confusing classes for each sample and therefore moves swiftly along the direction on the loss surface that seeks to resolve these ambiguities. The synergy between these two blocks helps the network to learn more effective feature representations. Comprehensive experiments are performed on five challenging datasets. Our approach outperforms existing methods using similar experimental setting on all five datasets.

In this paper, we present a novel Motion-Attentive Transition Network (MATNet) for zero-shot video object segmentation, which provides a new way of leveraging motion information to reinforce spatio-temporal object representation. An asymmetric attention block, called Motion-Attentive Transition (MAT), is designed within a two-stream encoder, which transforms appearance features into motion-attentive representations at each convolutional stage. In this way, the encoder becomes deeply interleaved, allowing for closely hierarchical interactions between object motion and appearance. This is superior to the typical two-stream architecture, which treats motion and appearance separately in each stream and often suffers from overfitting to appearance information. Additionally, a bridge network is proposed to obtain a compact, discriminative and scale-sensitive representation for multi-level encoder features, which is further fed into a decoder to achieve segmentation results. Extensive experiments on three challenging public benchmarks (i.e., DAVIS-16, FBMS and Youtube-Objects) show that our model achieves compelling performance against the state-of-the-arts. Code is available at: https://github.com/tfzhou/MATNet.

Human motion segmentation based on transfer subspace learning is a rising interest in action-related tasks. Although progress has been made, there are still several issues within the existing methods. First, existing methods transfer knowledge from source data to target tasks by learning domain-invariant features, but they ignore to preserve domain-specific knowledge. Second, the transfer subspace learning is employed in either low-level or high-level feature spaces, but few methods consider fusing multi-level features for subspace learning. To this end, we propose a novel multi-mutual consistency induced transfer subspace learning framework for human motion segmentation. Specifically, our model factorizes the source and target data into distinct multi-layer feature spaces and reduces the distribution gap between them through a multi-mutual consistency learning strategy. In this way, the domain-specific knowledge and domain-invariant properties can be explored simultaneously. Our model also conducts the transfer subspace learning on different layers to capture multi-level structural information. Further, to preserve the temporal correlations, we project the learned representations into a block-like space. The proposed model is efficiently optimized by using the Augmented Lagrange Multiplier (ALM) algorithm. Experimental results on four human motion datasets demonstrate the effectiveness of our method over other state-of-the-art approaches.

The key for fusion based hyperspectral image (HSI) super-resolution (SR) is to infer the posteriori of a latent HSI using appropriate image prior and likelihood that depends on degeneration. However, in practice the priors of high-dimensional HSIs can be extremely complicated and the degeneration is often unknown. Consequently most existing approaches that assume a shallow hand-crafted image prior and a pre-defined degeneration, fail to well generalize in real applications. To tackle this problem, we present an unsupervised adaptation learning (UAL) framework. Instead of directly modelling the complicated image prior, we propose to first implicitly learn a general image prior using deep networks and then adapt it to a specific HSI. Following this idea, we develop a two-stage SR network that leverages two consecutive modules: a fusion module and an adaptation module, to recover the latent HSI in a coarse-to-fine scheme. The fusion module is pretrained in a supervised manner on synthetic data to capture a spatial-spectral prior that is general across most HSIs. To adapt the learned general prior to the specific HSI under unknown degeneration, we introduce a simple degeneration network to assist learning both the adaptation module and the degeneration in an unsupervised way. In this way, the resultant image-specific prior and the estimated degeneration can benefit the inference of a more accurate posteriori, thereby increasing generalization capacity. To verify the efficacy of UAL, we extensively evaluate it on four benchmark datasets and report strong results that surpass existing approaches.

We propose a novel two-stage detection method, D2Det, that collectively addresses both precise localization and accurate classification. For precise localization, we introduce a dense local regression that predicts multiple dense box offsets for an object proposal. Different from traditional regression and keypoint-based localization employed in two-stage detectors, our dense local regression is not limited to a quantized set of keypoints within a fixed region and has the ability to regress position-sensitive real number dense offsets, leading to more precise localization. The dense local regression is further improved by a binary overlap prediction strategy that reduces the influence of background region on the final box regression. For accurate classification, we introduce a discriminative RoI pooling scheme that samples from various sub-regions of a proposal and performs adaptive weighting to obtain discriminative features. On MS COCO test-dev, our D2Det outperforms existing two-stage methods, with a single-model performance of 45.4 AP, using ResNet101 backbone. When using multi-scale training and inference, D2Det obtains AP of 50.1. In addition to detection, we adapt D2Det for instance segmentation, achieving a mask AP of 40.2 with a two-fold speedup, compared to the state-of-the-art. We also demonstrate the effectiveness of our D2Det on airborne sensors by performing experiments for object detection in UAV images (UAVDT dataset) and instance segmentation in satellite images (iSAID dataset). Source code is available at https://github.com/JialeCao001/D2Det.

We present a comprehensive study on a new task named camouflaged object detection (COD), which aims to identify objects that are “seamlessly” embedded in their surroundings. The high intrinsic similarities between the target object and the background make COD far more challenging than the traditional object detection task. To address this issue, we elaborately collect a novel dataset, called COD10K, which comprises 10,000 images covering camouflaged objects in various natural scenes, over 78 object categories. All the images are densely annotated with category, bounding-box, object-/instance-level, and matting-level labels. This dataset could serve as a catalyst for progressing many vision tasks, e.g., localization, segmentation, and alpha-matting, etc. In addition, we develop a simple but effective framework for COD, termed Search Identification Network (SINet). Without any bells and whistles, SINet outperforms various state-of-the-art object detection baselines on all datasets tested, making it a robust, general framework that can help facilitate future research in COD. Finally, we conduct a large-scale COD study, evaluating 13 cutting-edge models, providing some interesting findings, and showing several potential applications. Our research offers the community an opportunity to explore more in this new field. The code will be available at https://github.com/DengPingFan/SINet/.

Orthogonality is widely used for training deep neural networks (DNNs) due to its ability to maintain all singular values of the Jacobian close to 1 and reduce redundancy in representation. This paper proposes a computationally efficient and numerically stable orthogonalization method using Newton's iteration (ONI), to learn a layer-wise orthogonal weight matrix in DNNs. ONI works by iteratively stretching the singular values of a weight matrix towards 1. This property enables it to control the orthogonality of a weight matrix by its number of iterations. We show that our method improves the performance of image classification networks by effectively controlling the orthogonality to provide an optimal tradeoff between optimization benefits and representational capacity reduction. We also show that ONI stabilizes the training of generative adversarial networks (GANs) by maintaining the Lipschitz continuity of a network, similar to spectral normalization (SN), and further outperforms SN by providing controllable orthogonality.

Batch Normalization (BN) is extensively employed in various network architectures by performing standardization within mini-batches. A full understanding of the process has been a central target in the deep learning communities. Unlike existing works, which usually only analyze the standardization operation, this paper investigates the more general Batch Whitening (BW). Our work originates from the observation that while various whitening transformations equivalently improve the conditioning, they show significantly different behaviors in discriminative scenarios and training Generative Adversarial Networks (GANs). We attribute this phenomenon to the stochasticity that BW introduces. We quantitatively investigate the stochasticity of different whitening transformations and show that it correlates well with the optimization behaviors during training. We also investigate how stochasticity relates to the estimation of population statistics during inference. Based on our analysis, we provide a framework for designing and comparing BW algorithms in different scenarios. Our proposed BW algorithm improves the residual networks by a significant margin on ImageNet classification. Besides, we show that the stochasticity of BW can improve the GAN's performance with, however, the sacrifice of the training stability.

Several studies show that animal needs are often expressed through their faces. Though remarkable progress has been made towards the automatic understanding of human faces, this has not been the case with animal faces. There exists significant room for algorithmic advances that could realize automatic systems for interpreting animal faces. Besides scientific value, resulting technology will foster better and cheaper animal care. We believe the underlying research progress is mainly obstructed by the lack of an adequately annotated dataset of animal faces, covering a wide spectrum of animal species. To this end, we introduce a large-scale, hierarchical annotated dataset of animal faces, featuring 22.4K faces from 350 diverse species and 21 animal orders across biological taxonomy. These faces are captured `in-the-wild' conditions and are consistently annotated with 9 landmarks on key facial features. The dataset is structured and scalable by design; its development underwent four systematic stages involving rigorous, overall effort of over 6K man-hours. We benchmark it for face alignment using the existing art under two new problem settings. Results showcase its challenging nature, unique attributes and present definite prospects for novel, adaptive, and generalized face-oriented CV algorithms. Further benchmarking the dataset across face detection and fine-grained recognition tasks demonstrates its multi-task applications and room for improvement. The dataset is available at: https://fdmaproject.wordpress.com/.

Due to the advantages of real-time detection and improved performance, single-shot detectors have gained great attention recently. To solve the complex scale variations, single-shot detectors make scale-aware predictions based on multiple pyramid layers. However, the features in the pyramid are not scale-aware enough, which limits the detection performance. Two common problems in single-shot detectors caused by object scale variations can be observed: (1) small objects are easily missed; (2) the salient part of a large object is sometimes detected as an object. With this observation, we propose a new Neighbor Erasing and Transferring (NET) mechanism to reconfigure the pyramid features and explore scale-aware features. In NET, a Neighbor Erasing Module (NEM) is designed to erase the salient features of large objects and emphasize the features of small objects in shallow layers. A Neighbor Transferring Module (NTM) is introduced to transfer the erased features and highlight large objects in deep layers. With this mechanism, a single-shot network called NETNet is constructed for scale-aware object detection. In addition, we propose to aggregate nearest neighboring pyramid features to enhance our NET. NETNet achieves 38.5% AP at a speed of 27 FPS and 32.0% AP at a speed of 55 FPS on MS COCO dataset. As a result, NETNet achieves a better trade-off for real-time and accurate object detection.

Neural network pruning offers a promising prospect to facilitate deploying deep neural networks on resource-limited devices. However, existing methods are still challenged by the training inefficiency and labor cost in pruning designs, due to missing theoretical guidance of non-salient network components. In this paper, we propose a novel filter pruning method by exploring the High Rank of feature maps (HRank). Our HRank is inspired by the discovery that the average rank of multiple feature maps generated by a single filter is always the same, regardless of the number of image batches CNNs receive. Based on HRank, we develop a method that is mathematically formulated to prune filters with low-rank feature maps. The principle behind our pruning is that low-rank feature maps contain less information, and thus pruned results can be easily reproduced. Besides, we experimentally show that weights with high-rank feature maps contain more important information, such that even when a portion is not updated, very little damage would be done to the model performance. Without introducing any additional constraints, HRank leads to significant improvements over the state-of-the-arts in terms of FLOPs and parameters reduction, with similar accuracies. For example, with ResNet-110, we achieve a 58.2%-FLOPs reduction by removing 59.2% of the parameters, with only a small loss of 0.14% in top-1 accuracy on CIFAR-10. With Res-50, we achieve a 43.8%-FLOPs reduction by removing 36.7% of the parameters, with only a loss of 1.17% in the top-1 accuracy on ImageNet. The codes can be available at https://github.com/lmbxmu/HRank.

Medical image segmentation is important for computer aided diagnosis. Pixelwise manual annotations of large datasets require high expertise and is time consuming. Conventional data augmentations have limited benefit by not fully representing the underlying distribution of the training set, thus affecting model robustness when tested on images captured from different sources. Prior work leverages synthetic images for data augmentation ignoring the interleaved geometric relationship between different anatomical labels. We propose improvements over previous GAN-based medical image synthesis methods by jointly encoding the intrinsic relationship of geometry and shape. Latent space variable sampling results in diverse generated images from a base image and improves robustness. Augmented datasets using our method for automatic segmentation of retinal optical coherence tomography (OCT) images outperform existing methods on the public RETOUCH dataset having images captured from different acquisition procedures. Ablation studies and visual analysis also demonstrate benefits of integrating geometry and diversity.

Conventional unsupervised hashing methods usually take advantage of similarity graphs, which are either pre-computed in the high-dimensional space or obtained from random anchor points. On the one hand, existing methods uncouple the procedures of hash function learning and graph construction. On the other hand, graphs empirically built upon original data could introduce biased prior knowledge of data relevance, leading to sub-optimal retrieval performance. In this paper, we tackle the above problems by proposing an efficient and adaptive code-driven graph, which is updated by decoding in the context of an auto-encoder. Specifically, we introduce into our framework twin bottlenecks (i.e., latent variables) that exchange crucial information collaboratively. One bottleneck (i.e., binary codes) conveys the high-level intrinsic data structure captured by the code-driven graph to the other (i.e., continuous variables for low-level detail information), which in turn propagates the updated network feedback for the encoder to learn more discriminative binary codes. The auto-encoding learning objective literally rewards the code-driven graph to learn an optimal encoder. Moreover, the proposed model can be simply optimized by gradient descent without violating the binary constraints. Experiments on benchmarked datasets clearly show the superiority of our framework over the state-of-the-art hashing methods. Our source code can be found at https://github.com/ymcidence/TBH.

Human parsing is for pixel-wise human semantic understanding. As human bodies are underlying hierarchically structured, how to model human structures is the central theme in this task. Focusing on this, we seek to simultaneously exploit the representational capacity of deep graph networks and the hierarchical human structures. In particular, we provide following two contributions. First, three kinds of part relations, i.e., decomposition, composition, and dependency, are, for the first time, completely and precisely described by three distinct relation networks. This is in stark contrast to previous parsers, which only focus on a portion of the relations and adopt a type-agnostic relation modeling strategy. More expressive relation information can be captured by explicitly imposing the parameters in the relation networks to satisfy the specific characteristics of different relations. Second, previous parsers largely ignore the need for an approximation algorithm over the loopy human hierarchy, while we instead address an iterative reasoning process, by assimilating generic message-passing networks with their edge-typed, convolutional counterparts. With these efforts, our parser lays the foundation for more sophisticated and flexible human relation patterns of reasoning. Comprehensive experiments on five datasets demonstrate that our parser sets a new state-of-the-art on each.

Video-based person re-identification (re-ID) is an important research topic in computer vision. The key to tackling the challenging task is to exploit both spatial and temporal clues in video sequences. In this work, we propose a novel graph-based framework, namely Multi-Granular Hypergraph (MGH), to pursue better representational capabilities by modeling spatiotemporal dependencies in terms of multiple granularities. Specifically, hypergraphs with different spatial granularities are constructed using various levels of part-based features across the video sequence. In each hypergraph, different temporal granularities are captured by hyperedges that connect a set of graph nodes (i.e., part-based features) across different temporal ranges. Two critical issues (misalignment and occlusion) are explicitly addressed by the proposed hypergraph propagation and feature aggregation schemes. Finally, we further enhance the overall video representation by learning more diversified graph-level representations of multiple granularities based on mutual information minimization. Extensive experiments on three widely-adopted benchmarks clearly demonstrate the effectiveness of the proposed framework. Notably, 90.0% top-1 accuracy on MARS is achieved using MGH, outperforming the state-of-the-arts.

The availability of large-scale datasets has helped unleash the true potential of deep convolutional neural networks (CNNs). However, for the single-image denoising problem, capturing a real dataset is an unacceptably expensive and cumbersome procedure. Consequently, image denoising algorithms are mostly developed and evaluated on synthetic data that is usually generated with a widespread assumption of additive white Gaussian noise (AWGN). While the CNNs achieve impressive results on these synthetic datasets, they do not perform well when applied on real camera images, as reported in recent benchmark datasets. This is mainly because the AWGN is not adequate for modeling the real camera noise which is signal-dependent and heavily transformed by the camera imaging pipeline. In this paper, we present a framework that models camera imaging pipeline in forward and reverse directions. It allows us to produce any number of realistic image pairs for denoising both in RAW and sRGB spaces. By training a new image denoising network on realistic synthetic data, we achieve the state-of-the-art performance on real camera benchmark datasets. The parameters in our models are ~5 times lesser than the previous best method for RAW denoising. Furthermore, we demonstrate that the proposed framework generalizes beyond image denoising problem e.g., for color matching in stereoscopic cinema. The source code and pre-trained models are available at https://github.com/swz30/CycleISP.

Approximate Nearest Neighbor (ANN) search in high-dimensional space is a fundamental task in many applications. Locality-Sensitive Hashing (LSH) is a well-known methodology to solve the ANN problem with theoretical guarantees and empirical performance. We observe that existing LSH-based approaches target at the problem of designing search optimized indexes, which require a number of separate indexes and high index maintenance overhead, and hence impractical for high-dimensional streaming data processing. In this paper, we present PDA-LSH, a novel and practical disk-based LSH index that can offer efficient support for both updates and searches. Experiments on real-world datasets show that our proposal outperforms the state-of-the-art schemes by up to 10× on update performance and up to 2× on search performance.

One fundamental problem in deep learning is understanding the outstanding performance of deep Neural Networks (NNs) in practice. One explanation for the superiority of NNs is that they can realize a large class of complicated functions, i.e., they have powerful expressivity. The expressivity of a ReLU NN can be quantified by the maximal number of linear regions it can separate its input space into. In this paper, we provide several mathematical results needed for studying the linear regions of CNNs, and use them to derive the maximal and average numbers of linear regions for one-layer ReLU CNNs. Furthermore, we obtain upper and lower bounds for the number of linear regions of multi-layer ReLU CNNs. Our results suggest that deeper CNNs have more powerful expressivity than their shallow counterparts, while CNNs have more expressivity than fully-connected NNs per parameter.

We introduce kernels with random Fourier features in the meta-learning framework for few-shot learning. We propose meta variational random features (MetaVRF) to learn adaptive kernels for the base-learner, which is developed in a latent variable model by treating the random feature basis as the latent variable. We formulate the optimization of MetaVRF as a variational inference problem by deriving an evidence lower bound under the meta-learning framework. To incorporate shared knowledge from related tasks, we propose a context inference of the posterior, which is established by an LSTM architecture. The LSTM-based inference network can effectively integrate the context information of previous tasks with task-specific information, generating informative and adaptive features. The learned MetaVRF can produce kernels of high representational power with a relatively low spectral sampling rate and also enables fast adaptation to new tasks. Experimental results on a variety of few-shot regression and classification tasks demonstrate that MetaVRF delivers much better, or at least competitive, performance compared to existing meta-learning alternatives.

Residual networks (ResNets) represent a powerful type of convolutional neural network (CNN) architecture, widely adopted and used in various tasks. In this work we propose an improved version of ResNets. Our proposed improvements address all three main components of a ResNet: the flow of information through the network layers, the residual building block, and the projection shortcut. We are able to show consistent improvements in accuracy and learning convergence over the baseline. For instance, on ImageNet dataset, using the ResNet with 50 layers, for top-1 accuracy we can report a 1.19% improvement over the baseline in one setting and around 2% boost in another. Importantly, these improvements are obtained without increasing the model complexity. Our proposed approach allows us to train extremely deep networks, while the baseline shows severe optimization issues. We report results on three tasks over six datasets: image classification (ImageNet, CIFAR-10 and CIFAR-100), object detection (COCO) and video action recognition (Kinetics-400 and Something-Something-v2). In the deep learning era, we establish a new milestone for the depth of a CNN. We successfully train a 404-layer deep CNN on the ImageNet dataset and a 3002-layer network on CIFAR-10 and CIFAR-100, while the baseline is not able to converge at such extreme depths. Code and models are publicly available at: https://github.com/iduta/iresnet.

Image super-resolution (SR) and image inpainting are two topical problems in medical image processing. Existing methods for solving the problems are either tailored to recovering a high-resolution version of the low-resolution image or focus on filling missing values, thus inevitably giving rise to poor performance when the acquisitions suffer from multiple degradations. In this paper, we explore the possibility of super-resolving and inpainting images to handle multiple degradations and therefore improve their usability. We construct a unified and scalable framework to overcome the drawbacks of propagated errors caused by independent learning. We additionally provide improvements over previously proposed super-resolution approaches by modeling image degradation directly from data observations rather than bicubic downsampling. To this end, we propose HLH-GAN, which includes a high-to-low (H-L) GAN together with a low-to-high (L-H) GAN in a cyclic pipeline for solving the medical image degradation problem. Our comparative evaluation demonstrates that the effectiveness of the proposed method on different brain MRI datasets. In addition, our method outperforms many existing super-resolution and inpainting approaches.

Person re-identification (ReID) has recently received extensive research interests due to its diverse applications in multimedia analysis and computer vision. However, the majority of existing works focus on improving matching accuracy, while ignoring matching efficiency. In this work, we present a novel binary representation learning framework for efficient person ReID, namely Deep Local Binary Coding (DLBC). Different from existing deep binary ReID approaches, DLBC attempts to learn discriminative binary codes by explicitly interacting with local visual details. Specifically, DLBC first extracts a set of local features from spatially salient regions of pedestrian images. Subsequently, DLBC formulates a new binary-local semantic mutual information (BSMI) maximization term, based on which a self-lifting (SL) block is built to further exploit the semantic importance of local features. The BSMI term together with the SL block simultaneously enhances the dependency of binary codes on selected local features as well as their robustness to cross-view visual inconsistency. In addition, an efficient optimizing method is developed to train the proposed deep models with orthogonal and binary constraints. Extensive experiments reveal that DLBC significantly minimizes the accuracy gap between binary ReID methods and the state-of-the-art real-valued ones, whilst remarkably reducing query time and memory cost.

Occlusions, scale variation and numerous false positives still represent fundamental challenges in pedestrian detection. Intuitively, different sizes of receptive fields and more attention to the visible parts are required for detecting pedestrians with various scales and occlusion levels, respectively. However, these challenges have not been addressed well by existing pedestrian detectors. This paper presents a novel convolutional network, denoted as box guided convolution network (BGCNet), to tackle these challenges simultaneously in a unified framework. In particular, we proposed a box guided convolution (BGC) that can dynamically adjust the sizes of convolution kernels guided by the predicted bounding boxes. In this way, BGCNet provides position-aware receptive fields to address the challenge of large variations of scales. In addition, for the issue of heavy occlusion, the kernel parameters of BGC are spatially localized around the salient and mostly visible key points of a pedestrian, such as the head and foot, to effectively capture high-level semantic features to help detection. Furthermore, a local maximum (LM) loss is introduced to depress false positives and highlight true positives by forcing positives, rather than negatives, as local maximums, without any additional inference burden. We evaluate BGCNet on popular pedestrian detection benchmarks, and achieve the state-of-the-art results, with the significant performance improvement on heavily occluded and small-scale pedestrians.

Person re-identification has seen significant advancement in recent years. However, the ability of learned models to generalize to unknown target domains still remains limited. One possible reason for this is the lack of large-scale and diverse source training data, since manually labeling such a dataset is very expensive and privacy sensitive. To address this, we propose to automatically synthesize a large-scale person re-identification dataset following a set-up similar to real surveillance but with virtual environments, and then use the synthesized person images to train a generalizable person re-identification model. Specifically, we design a method to generate a large number of random UV texture maps and use them to create different 3D clothing models. Then, an automatic code is developed to randomly generate various different 3D characters with diverse clothes, races and attributes. Next, we simulate a number of different virtual environments using Unity3D, with customized camera networks similar to real surveillance systems, and import multiple 3D characters at the same time, with various movements and interactions along different paths through the camera networks. As a result, we obtain a virtual dataset, called RandPerson, with 1,801,816 person images of 8,000 identities. By training person re-identification models on these synthesized person images, we demonstrate, for the first time, that models trained on virtual data can generalize well to unseen target images, surpassing the models trained on various real-world datasets, including CUHK03, Market-1501, DukeMTMC-reID, and almost MSMT17. The RandPerson dataset is available at https://github.com/VideoObjectSearch/RandPerson.

In this paper, we propose a cross-modal network architecture search (NAS) algorithm for VQA, termed as k-Armed Bandit based NAS (KAB-NAS). KAB-NAS regards the design of each layer as a k-armed bandit problem and updates the preference of each candidate via numerous samplings in a single-shot search framework. To establish an effective search space, we further propose a new architecture termed Automatic Graph Attention Network (AGAN), and extend the popular self-attention layer with three graph structures, denoted as dense-graph, co-graph and separate-graph.These graph layers are used to form the direction of information propagation in the graph network, and their optimal combinations are searched by KAB-NAS. To evaluate KAB-NAS and AGAN, we conduct extensive experiments on two VQA benchmark datasets, i.e., VQA2.0 and GQA, and also test AGAN with the popular BERT-style pre-training. The experimental results show that with the help of KAB-NAS, AGAN can achieve the state-of-the-art performance on both benchmark datasets with much fewer parameters and computations.

This paper proposes a human parsing based texture transfer model via cross-view consistency learning to generate the texture of 3D human body from a single image. We use the semantic parsing of human body as input for providing both the shape and pose information to reduce the appearance variation of human image and preserve the spatial distribution of semantic parts. Meanwhile, in order to improve the prediction for textures of invisible parts, we explicitly enforce the consistency across different views of the same subject by exchanging the textures predicted by two views to render images during training. The perception loss and total variation regularization are optimized to maximize the similarity between rendered and input images, which does not necessitate extra 3D texture supervision. Experimental results on pedestrian images and fashion photos demonstrate that our method can produce higher quality textures with convincing details than other texture generation methods.

In this paper, we introduce variational semantic memory into meta-learning to acquire long-term knowledge for few-shot learning. The variational semantic memory accrues and stores semantic information for the probabilistic inference of class prototypes in a hierarchical Bayesian framework. The semantic memory is grown from scratch and gradually consolidated by absorbing information from tasks it experiences. By doing so, it is able to accumulate long-term, general knowledge that enables it to learn new concepts of objects. We formulate memory recall as the variational inference of a latent memory variable from addressed contents, which offers a principled way to adapt the knowledge to individual tasks. Our variational semantic memory, as a new long-term memory module, confers principled recall and update mechanisms that enable semantic information to be efficiently accrued and adapted for few-shot learning. Experiments demonstrate that the probabilistic modelling of prototypes achieves a more informative representation of object classes compared to deterministic vectors. The consistent new state-of-the-art performance on four benchmarks shows the benefit of variational semantic memory in boosting few-shot recognition.

Recent neural architectures in named entity recognition (NER) have yielded state-of-the-art performance on single domain data such as newswires. However, they still suffer from (i) requiring massive amounts of training data to avoid overfitting; (ii) huge performance degradation when there is a domain shift in the data distribution between training and testing. In this paper, we investigate the problem of domain adaptation for NER under homogeneous and heterogeneous settings. We propose MetaNER, a novel meta-learning approach for domain adaptation in NER. Specifically, MetaNER incorporates meta-learning and adversarial training strategies to encourage robust, general and transferable representations for sequence labeling. The key advantage of MetaNER is that it is capable of adapting to new unseen domains with a small amount of annotated data from those domains. We extensively evaluate MetaNER on multiple datasets under homogeneous and heterogeneous settings. The experimental results show that MetaNER achieves state-of-the-art performance against eight baselines. Impressively, MetaNER surpasses the in-domain performance using only 16.17% and 34.76% of target domain data on average for homogeneous and heterogeneous settings, respectively.

Abstract While many image colorization algorithms have recently shown the capability of producing plausible color versions from gray-scale photographs, they still suffer from limited semantic understanding. To address this shortcoming, we propose to exploit pixelated object semantics to guide image colorization. The rationale is that human beings perceive and distinguish colors based on the semantic categories of objects. Starting from an autoregressive model, we generate image color distributions, from which diverse colored results are sampled. We propose two ways to incorporate object semantics into the colorization model: through a pixelated semantic embedding and a pixelated semantic generator. Specifically, the proposed network includes two branches. One branch learns what the object is, while the other branch learns the object colors. The network jointly optimizes a color embedding loss, a semantic segmentation loss and a color generation loss, in an end-to-end fashion. Experiments on Pascal VOC2012 and COCO-stuff reveal that our network, when trained with semantic segmentation labels, produces more realistic and finer results compared to the colorization state-of-the-art.

Compressive sensing (CS)-based image coding scheme has been enthusiastically studied, but it still has a poor rate-distortion performance compared with the traditional image coding techniques. In this paper, we propose a CS multi-layer residual coding scheme to rectify this problem to a certain extent. By dividing CS measurements into multi-layers and predicting a particular layer's measurements with all its preceding layers' measurements, we can transform CS measurements into multi-layer residual coefficients, which are easier to compress. By calculating the residual between the quantized ground-truth CS measurements and their corresponding quantized inference measurements and using Huffman coding to associate each residual quantization index with a binary code, we can reduce the redundancies among CS measurements efficiently. Besides, the prediction and quantization process is designed to be layer-independent, which can save much of the encoding time. The proposed approach introduces a novel framework for using CS in the compression domain. The experimental results show that the proposed scheme can significantly outperform JPEG2000 and approach or reach the performance of HEVC-Intra on some test images.

In this paper, we propose a computationally efficient and consistently accurate spatiotemporal salient object detection method to identify the most noticeable object in a video sequence. Intuitively, the underlying motion in a video is a more stable saliency indicator than the apparent color cues that often contain significant variations and complex structures. Based on this observation, we build an efficient and accurate spatiotemporal saliency detection method that uses motion information as a leverage to locate the most dynamic regions in a video sequence. We first analyze the optical flow field to obtain foreground priors, and then incorporate spatial saliency features such as appearance contrasts and compactness measures, into a multi-cue integration framework to combine various saliency cues and achieve temporal consistency. Rigorous experiments on the challenging SegTrackV1, SegTrackV2, and FBMS datasets demonstrate that our method generates comparable or superior performance to state-of-the-art methods while running almost 100× faster at only 0.08 sec/frame. Promising performance and rapid speed imply that the proposed spatiotemporal saliency method can be easily involved in various vision applications.

Sketch-based image retrieval (SBIR) is a challenging task due to the large cross-domain gap between sketches and natural images. How to align abstract sketches and natural images into a common high-level semantic space remains a key problem in SBIR. In this paper, we propose a novel semi-heterogeneous three-way joint embedding network (Semi3-Net), which integrates three branches (a sketch branch, a natural image branch, and an edgemap branch) to learn more discriminative cross-domain feature representations for the SBIR task. The key insight lies with how we cultivate the mutual and subtle relationships amongst the sketches, natural images, and edgemaps. A semi-heterogeneous feature mapping is designed to extract bottom features from each domain, where the sketch and edgemap branches are shared while the natural image branch is heterogeneous to the other branches. In addition, a joint semantic embedding is introduced to embed the features from different domains into a common high-level semantic space, where all of the three branches are shared. To further capture informative features common to both natural images and the corresponding edgemaps, a co-attention model is introduced to conduct common channel-wise feature recalibration between different domains. A hybrid-loss mechanism is designed to align the three branches, where an alignment loss and a sketch-edgemap contrastive loss are presented to encourage the network to learn invariant cross-domain representations. Experimental results on two widely used category-level datasets (Sketchy and TU-Berlin Extension) demonstrate that the proposed method outperforms state-of-the-art methods.

In this paper, we propose a novel superpixel segmentation method using an adaptive nonlocal random walk (ANRW) algorithm. There are three main steps in our image superpixel segmentation algorithm. Our method is based on the random walk model, in which the seed points are produced to generate the initial superpixels by a gradient-based method in the first step. In the second step, the ANRW is proposed to get the initial superpixels by adjusting the NRW to obtain a better image and superpixel segmentation. In the last step, these small superpixels are merged to get the final regular and compact superpixels. The experimental results demonstrate that our method achieves a better superpixel performance than the state-of-the-art methods. Our source code will be available at: http://github.com/shenjianbing/ANRW.

Previous methods on text image motion deblurring seldom consider the sparse characteristics of the blur kernel. This paper proposes a new text image motion deblurring method by exploiting the sparse properties of both text image itself and kernel. It incorporates the L0-norm for regularizing the blur kernel in the deblurring model, besides the L0sparse priors for the text image and its gradient. Such a L0-norm-based model is efficiently optimized by half-quadratic splitting coupled with the fast conjugate descent method. To further improve the quality of the recovered kernel, a structure-preserving kernel denoising method is also developed to filter out the noisy pixels, yielding a clean kernel curve. Experimental results show the superiority of the proposed method. The source code and results are available at: https://github.com/shenjianbing/text-image-deblur.

Visual tracking addresses the problem of localizing an arbitrary target in video according to the annotated bounding box. In this article, we present a novel tracking method by introducing the attention mechanism into the Siamese network to increase its matching discrimination. We propose a new way to compute attention weights to improve matching performance by a sub-Siamese network [Attention Net (A-Net)], which locates attentive parts for solving the searching problem. In addition, features in higher layers can preserve more semantic information while features in lower layers preserve more location information. Thus, in order to solve the tracking failure cases by the higher layer features, we fully utilize location and semantic information by multilevel features and propose a new way to fuse multiscale response maps from each layer to obtain a more accurate position estimation of the object. We further propose a hierarchical attention Siamese network by combining the attention weights and multilayer integration for tracking. Our method is implemented with a pretrained network which can outperform most well-trained Siamese trackers even without any fine-tuning and online updating. The comparison results with the state-of-the-art methods on popular tracking benchmarks show that our method achieves better performance. Our source code and results will be available at https://github.com/shenjianbing/HASN.

Video action understanding is an active field of intelligent video analytics, and contextual information in the videos has gained lots of attention for better action understanding. However, most existing works focus on using contextual information for supervised or semi-supervised analysis, and how to effectively use contextual information to boost the unsupervised action clustering performance is still a challenging problem. In this article, we propose a recursive constrained framework for unsupervised video action clustering by utilizing the contextual information of the action and scene. Considering the unique contextual characteristics of video action, action context clustering solution and scene context clustering solution are obtained simultaneously. Based on these two solutions, a recursive priori propagation is proposed to exploit information gain of the priori clustering solutions, and then the information gain is fed back into the procedures of both subspace representation and spectral clustering. Specifically, to explore the unknown relationships in the priori clustering solutions, the constraint-guided subspace representation is introduced by fusing the recursive priori constraint into the self-representation model. Taking priori information and multiview features into consideration, the priori-inherited multiview spectral clustering is proposed to obtain more discriminative spectral embeddings for action clustering. Experiments on three video benchmark datasets demonstrate that the proposed method outperforms state-of-the-art methods.

Image deraining is an important yet challenging image processing task. Though deterministic image deraining methods are developed with encouraging performance, they are infeasible to learn flexible representations for probabilistic inference and diverse predictions. Besides, rain intensity varies both in spatial locations and across color channels, making this task more difficult. In this paper, we propose a Conditional Variational Image Deraining (CVID) network for better deraining performance, leveraging the exclusive generative ability of Conditional Variational Auto-Encoder (CVAE) on providing diverse predictions for the rainy image. To perform spatially adaptive deraining, we propose a spatial density estimation (SDE) module to estimate a rain density map for each image. Since rain density varies across different color channels, we also propose a channel-wise (CW) deraining scheme. Experiments on synthesized and real-world datasets show that the proposed CVID network achieves much better performance than previous deterministic methods on image deraining. Extensive ablation studies validate the effectiveness of the proposed SDE module and CW scheme in our CVID network. The code is available at https://github.com/Yingjun-Du/VID.

Zero-shot learning (ZSL) is a challenging task due to the lack of unseen class data during training. Existing works attempt to establish a mapping between the visual and class spaces through a common intermediate semantic space. The main limitation of existing methods is the strong bias towards seen class, known as the domain shift problem, which leads to unsatisfactory performance in both conventional and generalized ZSL tasks. To tackle this challenge, we propose to convert ZSL to the conventional supervised learning by generating features for unseen classes. To this end, a joint generative model that couples variational autoencoder (VAE) and generative adversarial network (GAN), called Zero-VAE-GAN, is proposed to generate high-quality unseen features. To enhance the class-level discriminability, an adversarial categorization network is incorporated into the joint framework. Besides, we propose two self-training strategies to augment unlabeled unseen features for the transductive extension of our model, addressing the domain shift problem to a large extent. Experimental results on five standard benchmarks and a large-scale dataset demonstrate the superiority of our generative model over the state-of-the-art methods for conventional, especially generalized ZSL tasks. Moreover, the further improvement of the transductive setting demonstrates the effectiveness of the proposed self-training strategies.

Non-local self similarity (NSS) is a powerful prior of natural images for image denoising. Most of existing denoising methods employ similar patches, which is a patch-level NSS prior. In this paper, we take one step forward by introducing a pixel-level NSS prior, i.e., searching similar pixels across a non-local region. This is motivated by the fact that finding closely similar pixels is more feasible than similar patches in natural images, which can be used to enhance image denoising performance. With the introduced pixel-level NSS prior, we propose an accurate noise level estimation method, and then develop a blind image denoising method based on the lifting Haar transform and Wiener filtering techniques. Experiments on benchmark datasets demonstrate that, the proposed method achieves much better performance than previous non-deep methods, and is still competitive with existing state-of-the-art deep learning based methods on real-world image denoising. The code is publicly available athttps://github.com/njusthyk1972/NLH.

The ability to synthesize multi-modality data is highly desirable for many computer-aided medical applications, e.g. clinical diagnosis and neuroscience research, since rich imaging cohorts offer diverse and complementary information unraveling human tissues. However, collecting acquisitions can be limited by adversary factors such as patient discomfort, expensive cost and scanner unavailability. In this paper, we propose a multi-task coherent modality transferable GAN (MCMT-GAN) to address this issue for brain MRI synthesis in an unsupervised manner. Through combining the bidirectional adversarial loss, cycle-consistency loss, domain adapted loss and manifold regularization in a volumetric space, MCMT-GAN is robust for multi-modality brain image synthesis with visually high fidelity. In addition, we complement discriminators collaboratively working with segmentors which ensure the usefulness of our results to segmentation task. Experiments evaluated on various cross-modality synthesis show that our method produces visually impressive results with substitutability for clinical post-processing and also exceeds the state-of-the-art methods.

The fundamental problem of Zero-Shot Learning (ZSL) is that the one-hot label space is discrete, which leads to a complete loss of the relationships between seen and unseen classes. Conventional approaches rely on using semantic auxiliary information, e.g. attributes, to re-encode each class so as to preserve the inter-class associations. However, existing learning algorithms only focus on unifying visual and semantic spaces without jointly considering the label space. More importantly, because the final classification is conducted in the label space through a compatibility function, the gap between attribute and label spaces leads to significant performance degradation. Therefore, this paper proposes a novel pathway that uses the label space to jointly reconcile visual and semantic spaces directly, which is named Attributing Label Space (ALS). In the training phase, one-hot labels of seen classes are directly used as prototypes in a common space, where both images and attributes are mapped. Since mappings can be optimized independently, the computational complexity is extremely low. In addition, the correlation between semantic attributes has less influence on visual embedding training because features are mapped into labels instead of attributes. In the testing phase, the discrete condition of label space is removed, and priori one-hot labels are used to denote seen classes and further compose labels of unseen classes. Therefore, the label space is very discriminative for the Generalized ZSL (GZSL), which is more reasonable and challenging for real-world applications. Extensive experiments on five benchmarks manifest improved performance over all of compared state-of-the-art methods.

Convolutional Neural Networks have achieved significant success across multiple computer vision tasks. However, they are vulnerable to carefully crafted, human-imperceptible adversarial noise patterns which constrain their deployment in critical security-sensitive systems. This paper proposes a computationally efficient image enhancement approach that provides a strong defense mechanism to effectively mitigate the effect of such adversarial perturbations. We show that deep image restoration networks learn mapping functions that can bring off-the-manifold adversarial samples onto the natural image manifold, thus restoring classification towards correct classes. A distinguishing feature of our approach is that, in addition to providing robustness against attacks, it simultaneously enhances image quality and retains models performance on clean images. Furthermore, the proposed method does not modify the classifier or requires a separate mechanism to detect adversarial images. The effectiveness of the scheme has been demonstrated through extensive experiments, where it has proven a strong defense in gray-box settings. The proposed scheme is simple and has the following advantages: (1) it does not require any model training or parameter optimization, (2) it complements other existing defense mechanisms, (3) it is agnostic to the attacked model and attack type and (4) it provides superior performance across all popular attack algorithms. Our codes are publicly available at https://github.com/aamir-mustafa/super-resolution-adversarial-defense.

Recognising remote sensing scene images remains challenging due to large visual-semantic discrepancies. These mainly arise due to the lack of detailed annotations that can be employed to align pixel-level representations with high-level semantic labels. As the tagging process is labour-intensive and subjective, we hereby propose a novel Multi-Granularity Canonical Appearance Pooling (MG-CAP) to automatically capture the latent ontological structure of remote sensing datasets. We design a granular framework that allows progressively cropping the input image to learn multi-grained features. For each specific granularity, we discover the canonical appearance from a set of pre-defined transformations and learn the corresponding CNN features through a maxout-based Siamese style architecture. Then, we replace the standard CNN features with Gaussian covariance matrices and adopt the proper matrix normalisations for improving the discriminative power of features. Besides, we provide a stable solution for training the eigenvalue-decomposition function (EIG) in a GPU and demonstrate the corresponding back-propagation using matrix calculus. Extensive experiments have shown that our framework can achieve promising results in public remote sensing scene datasets.

Video-based person re-identification (re-ID) refers to matching people across camera views from arbitrary unaligned video footages. Existing methods rely on supervision signals to optimise a projected space under which the distances between inter/intra-videos are maximised/minimised. However, this demands exhaustively labelling people across camera views, rendering them unable to be scaled in large networked cameras. Also, it is noticed that learning effective video representations with view invariance is not explicitly addressed for which features exhibit different distributions otherwise. Thus, matching videos for person re-ID demands flexible models to capture the dynamics in time-series observations and learn view-invariant representations with access to limited labeled training samples. In this paper, we propose a novel few-shot deep learning approach to videobased person re-ID, to learn comparable representations that are discriminative and view-invariant. The proposed method is developed on the variational recurrent neural networks (VRNNs) and trained adversarially to produce latent variables with temporal dependencies that are highly discriminative yet view-invariant in matching persons. Through extensive experiments conducted on three benchmark datasets, we empirically show the capability of our method in creating view-invariant temporal features and state-of-the-art performance achieved by our method.

Supervised deep networks have achieved promising performance on image denoising, by learning image priors and noise statistics on plenty pairs of noisy and clean images. Unsupervised denoising networks are trained with only noisy images. However, for an unseen corrupted image, both supervised and unsupervised networks ignore either its particular image prior, the noise statistics, or both. That is, the networks learned from external images inherently suffer from a domain gap problem: the image priors and noise statistics are very different between the training and test images. This problem becomes more clear when dealing with the signal dependent realistic noise. To circumvent this problem, in this work, we propose a novel "Noisy-As-Clean" (NAC) strategy of training self-supervised denoising networks. Specifically, the corrupted test image is directly taken as the "clean" target, while the inputs are synthetic images consisted of this corrupted image and a second yet similar corruption. A simple but useful observation on our NAC is: as long as the noise is weak, it is feasible to learn a self-supervised network only with the corrupted image, approximating the optimal parameters of a supervised network learned with pairs of noisy and clean images. Experiments on synthetic and realistic noise removal demonstrate that, the DnCNN and ResNet networks trained with our self-supervised NAC strategy achieve comparable or better performance than the original ones and previous supervised/unsupervised/self-supervised networks. The code is publicly available at https://github.com/csjunxu/Noisy-As-Clean.

Retinex theory is developed mainly to decompose an image into the illumination and reflectance components by analyzing local image derivatives. In this theory, larger derivatives are attributed to the changes in reflectance, while smaller derivatives are emerged in the smooth illumination. In this paper, we utilize exponentiated local derivatives (with an exponent γ) of an observed image to generate its structure map and texture map. The structure map is produced by been amplified with γ > 1, while the texture map is generated by been shrank with γ < 1. To this end, we design exponential filters for the local derivatives, and present their capability on extracting accurate structure and texture maps, influenced by the choices of exponents γ. The extracted structure and texture maps are employed to regularize the illumination and reflectance components in Retinex decomposition. A novel Structure and Texture Aware Retinex (STAR) model is further proposed for illumination and reflectance decomposition of a single image. We solve the STAR model by an alternating optimization algorithm. Each sub-problem is transformed into a vectorized least squares regression, with closed-form solutions. Comprehensive experiments on commonly tested datasets demonstrate that, the proposed STAR model produce better quantitative and qualitative performance than previous competing methods, on illumination and reflectance decomposition, low-light image enhancement, and color correction. The code is publicly available at https://github.com/csjunxu/STAR.

Coronavirus Disease 2019 (COVID-19) spread globally in early 2020, causing the world to face an existential health crisis. Automated detection of lung infections from computed tomography (CT) images offers a great potential to augment the traditional healthcare strategy for tackling COVID-19. However, segmenting infected regions from CT slices faces several challenges, including high variation in infection characteristics, and low intensity contrast between infections and normal tissues. Further, collecting a large amount of data is impractical within a short time period, inhibiting the training of a deep model. To address these challenges, a novel COVID-19 Lung Infection Segmentation Deep Network (Inf-Net) is proposed to automatically identify infected regions from chest CT slices. In our Inf-Net, a parallel partial decoder is used to aggregate the high-level features and generate a global map. Then, the implicit reverse attention and explicit edge-attention are utilized to model the boundaries and enhance the representations. Moreover, to alleviate the shortage of labeled data, we present a semi-supervised segmentation framework based on a randomly selected propagation strategy, which only requires a few labeled images and leverages primarily unlabeled data. Our semi-supervised framework can improve the learning ability and achieve a higher performance. Extensive experiments on our COVID-SemiSeg and real CT volumes demonstrate that the proposed Inf-Net outperforms most cutting-edge segmentation models and advances the state-of-the-art performance.

Magnetic resonance imaging (MRI) is a widely used neuroimaging technique that can provide images of different contrasts (i.e., modalities). Fusing this multi-modal data has proven particularly effective for boosting model performance in many tasks. However, due to poor data quality and frequent patient dropout, collecting all modalities for every patient remains a challenge. Medical image synthesis has been proposed as an effective solution, where any missing modalities are synthesized from the existing ones. In this paper, we propose a novel Hybrid-fusion Network (Hi-Net) for multi-modal MR image synthesis, which learns a mapping from multi-modal source images (i.e., existing modalities) to target images (i.e., missing modalities). In our Hi-Net, a modality-specific network is utilized to learn representations for each individual modality, and a fusion network is employed to learn the common latent representation of multi-modal data. Then, a multi-modal synthesis network is designed to densely combine the latent representation with hierarchical features from each modality, acting as a generator to synthesize the target images. Moreover, a layer-wise multi-modal fusion strategy effectively exploits the correlations among multiple modalities, where a Mixed Fusion Block (MFB) is proposed to adaptively weight different fusion strategies. Extensive experiments demonstrate the proposed model outperforms other state-of-the-art medical image synthesis methods.

Human Action Recognition (HAR) for CCTV-oriented applications is still a challenging problem. Real-world scenarios HAR implementations is difficult because of the gap between Deep Learning data requirements and what the CCTV-based frameworks can offer in terms of data recording equipments. We propose to reduce this gap by exploiting human poses provided by the OpenPose, which has been already proven to be an effective detector in CCTV-like recordings for tracking applications. Therefore, in this work, we first propose ActionXPose: a novel 2D pose-based approach for pose-level HAR. ActionXPose extracts low- and high-level features from body poses which are provided to a Long Short-Term Memory Neural Network and a 1D Convolutional Neural Network for the classification. We also provide a new dataset, named ISLD, for realistic pose-level HAR in a CCTV-like environment, recorded in the Intelligent Sensing Lab. ActionXPose is extensively tested on ISLD under multiple experimental settings, e.g. Dataset Augmentation and Cross-Dataset setting, as well as revising other existing datasets for HAR. ActionXPose achieves state-of-the-art performance in terms of accuracy, very high robustness to occlusions and missing data, and promising results for practical implementation in real-world applications.

Accurately categorizing Flickr images into multiple pre-defined communities (e.g., “architecture” and “peaceful”) is an indispensable technique in multimedia analysis, graphic design, fashion recommendation, etc. In practice, these communities are constructed and updated manually, which is subjective and intolerably time consuming. To alleviate these shortcomings, a noise-refined deep matrix factorization (MF) framework is proposed to intelligently discover communities from million-scale Flickr users, wherein the semantic tag correlations and community correlations are simultaneously encoded. More specifically, it is believable that Flickr communities are high-level clues on the basis of human visual semantic perception. Thereby, a MF algorithm is employed to approximate the community label matrix by the product of pairwise factor matrices, which represent the latent representations of user-provided tags and the corresponding basis matrix respectively. Subsequently, an end-to-end deep model is formulated to hierarchically derive the latent deep representation from raw image pixels to semantic tags. To robustly handle contaminated image semantic tags and community labels, an l1norm constraint is encoded to enhance the MF. Meanwhile, to optimally exploit the rich context information of Flickr images, the intrinsic structure between image semantic tags and between communities are collaboratively captured. Finally, the upgraded MF and the deep model are seamlessly combined into a unified framework, which is solved by an iterative algorithm. Experiments on 2 M Flickr images have demonstrated the superiority of our approach. Besides, the discovered Flickr communities can improve photo retargeting and visual aesthetics assessment significantly.

Kernel selection is of fundamental importance for the generalization of kernel methods. This article proposes an approximate approach for kernel selection by exploiting the approximability of kernel selection and the computational virtue of kernel matrix approximation. We define approximate consistency to measure the approximability of the kernel selection problem. Based on the analysis of approximate consistency, we solve the theoretical problem of whether, under what conditions, and at what speed, the approximate criterion is close to the accurate one, establishing the foundations of approximate kernel selection. We introduce two selection criteria based on error estimation and prove the approximate consistency of the multilevel circulant matrix (MCM) approximation and Nyström approximation under these criteria. Under the theoretical guarantees of the approximate consistency, we design approximate algorithms for kernel selection, which exploits the computational advantages of the MCM and Nyström approximations to conduct kernel selection in a linear or quasi-linear complexity. We experimentally validate the theoretical results for the approximate consistency and evaluate the effectiveness of the proposed kernel selection algorithms.

Both interclass variances and intraclass similarities are crucial for improving the classification performance of discriminative dictionary learning (DDL) algorithms. However, existing DDL methods often ignore the combination between the interclass and intraclass properties of dictionary atoms and coding coefficients. To address this problem, in this paper, we propose a discriminative Fisher embedding dictionary learning (DFEDL) algorithm that simultaneously establishes Fisher embedding models on learned atoms and coefficients. Specifically, we first construct a discriminative Fisher atom embedding model by exploring the Fisher criterion of the atoms, which encourages the atoms of the same class to reconstruct the corresponding training samples as much as possible. At the same time, a discriminative Fisher coefficient embedding model is formulated by imposing the Fisher criterion on the profiles (row vectors of the coding coefficient matrix) and coding coefficients, which forces the coding coefficient matrix to become a block-diagonal matrix. Since the profiles can indicate which training samples are represented by the corresponding atoms, the proposed two discriminative Fisher embedding models can alternatively and interactively promote the discriminative capabilities of the learned dictionary and coding coefficients. The extensive experimental results demonstrate that the proposed DFEDL algorithm achieves superior performance in comparison with some state-of-the-art dictionary learning algorithms on both hand-crafted and deep learning-based features.

Feature representation learning, an emerging topic in recent years, has achieved great progress. Powerful learned features can lead to excellent classification accuracy. In this article, a selective and robust feature representation framework with a supervised constraint (SRSC) is presented. SRSC seeks a selective, robust, and discriminative subspace by transforming the original feature space into the category space. Particularly, we add a selective constraint to the transformation matrix (or classifier parameter) that can select discriminative dimensions of the input samples. Moreover, a supervised regularization is tailored to further enhance the discriminability of the subspace. To relax the hard zero-one label matrix in the category space, an additional error term is also incorporated into the framework, which can lead to a more robust transformation matrix. SRSC is formulated as a constrained least square learning (feature transforming) problem. For the SRSC problem, an inexact augmented Lagrange multiplier method (ALM) is utilized to solve it. Extensive experiments on several benchmark data sets adequately demonstrate the effectiveness and superiority of the proposed method. The proposed SRSC approach has achieved better performances than the compared counterpart methods.

Zero-shot learning (ZSL), a type of structured multioutput learning, has attracted much attention due to its requirement of no training data for target classes. Conventional ZSL methods usually project visual features into semantic space and assign labels by finding their nearest prototypes. However, this type of nearest neighbor search (NNS)-based method often suffers from great performance degradation because of the nonuniform variances between different categories. In this article, we propose a probabilistic framework by taking covariance into account to deal with the above-mentioned problem. In this framework, we define a new latent space, which has two characteristics. The first is that the features in this space should gather within the classes and scatter between the classes, which is implemented by triplet learning; the second is that the prototypes of unseen classes are synthesized with nonnegative coefficients, which are generated by nonnegative matrix factorization (NMF) of relations between the seen classes and the unseen classes in attribute space. During training, the learned parameters are the projection model for triplet network and the nonnegative coefficients between the unseen classes and the seen classes. In the testing phase, visual features are projected into latent space and assigned with the labels that have the maximum probability among unseen classes for classic ZSL or within all classes for generalized ZSL. Extensive experiments are conducted on four popular data sets, and the results show that the proposed method can outperform the state-of-the-art methods in most circumstances.

Despite the competitive prediction performance, recent deep image quality models suffer from the following limitations. First, it is deficiently effective to interpret and quantify the region-level quality, which contributes to global features during deep architecture training. Second, human visual perception is sensitive to compositional features (i.e., the sophisticated spatial configurations among regions), but explicitly incorporating them into a deep model is challenging. Third, the state-of-the-art deep quality models typically use rectangular image patches as inputs, but there is no evidence that these rectangles can reflect arbitrarily shaped objects, such as beaches and jungles. By defining the complet, which is a set of image segments collaboratively characterizing the spatial/geometric distribution of multiple visual elements, we propose a novel quality-modeling framework that involves two key modules: a complet ranking algorithm and a spatially-aware dual aggregation network (SDA-Net). Specifically, to describe the region-level quality features, we build complets to characterize the high-order spatial interactions among the arbitrarily shaped segments in each image. To obtain complets that are highly descriptive to image compositions, a weakly supervised complet ranking algorithm is designed by quantifying the quality of each complet. The algorithm seamlessly encodes three factors: the image-level quality discrimination, weakly supervised constraint, and complet geometry of each image. Based on the top-ranking complets, a novel multi-column convolutional neural network (CNN) called SDA-Net is designed, which supports input segments with arbitrary shapes. The key is a dual-aggregation mechanism that fuses the intracomplet deep features and the intercomplet deep features under a unified framework. Thorough experimental validations on a series of benchmark data sets demonstrated the superiority of our method.

We propose a framework for analyzing and comparing distributions without imposing any parametric assumptions via empirical likelihood methods. Our framework is used to study two fundamental statistical test problems: the two-sample test and the goodness-of-fit test. For the two-sample test, we need to determine whether two groups of samples are from different distributions; for the goodness-of-fit test, we examine how likely it is that a set of samples is generated from a known target distribution. Specifically, we propose empirical likelihood ratio (ELR) statistics for the two-sample test and the goodness-of-fit test, both of which are of linear time complexity and show higher power (i.e., the probability of correctly rejecting the null hypothesis) than the existing linear statistics for high-dimensional data. We prove the nonparametric Wilks’ theorems for the ELR statistics, which illustrate that the limiting distributions of the proposed ELR statistics are chi-square distributions. With these limiting distributions, we can avoid bootstraps or simulations to determine the threshold for rejecting the null hypothesis, which makes the ELR statistics more efficient than the recently proposed linear statistic, finite set Stein discrepancy (FSSD). We also prove the consistency of the ELR statistics, which guarantees that the test power goes to 1 as the number of samples goes to infinity. In addition, we experimentally demonstrate and theoretically analyze that FSSD has poor performance or even fails to test for high-dimensional data. Finally, we conduct a series of experiments to evaluate the performance of our ELR statistics as compared to state-of-the-art linear statistics.

Kernel selection is fundamental to the generalization performance of kernel-based learning algorithms. Approximate kernel selection is an efficient kernel selection approach that exploits the convergence property of the kernel selection criteria and the computational virtue of kernel matrix approximation. The convergence property is measured by the notion of approximate consistency. For the existing Nyström approximations, whose sampling distributions are independent of the specific learning task at hand, it is difficult to establish the strong approximate consistency. They mainly focus on the quality of the low-rank matrix approximation, rather than the performance of the kernel selection criterion used in conjunction with the approximate matrix. In this paper, we propose a novel Nyström approximate kernel selection algorithm by customizing a criterion-driven adaptive sampling distribution for the Nyström approximation, which adaptively reduces the error between the approximate and accurate criteria. We theoretically derive the strong approximate consistency of the proposed Nyström approximate kernel selection algorithm. Finally, we empirically evaluate the approximate consistency of our algorithm as compared to state-of-the-art methods.

Sketch-based 3D shape retrieval has been extensively studied in recent works, most of which focus on improving the retrieval accuracy, whilst neglecting the efficiency. In this paper, we propose a novel framework for efficient sketch-based 3D shape retrieval, i.e., Deep Sketch-Shape Hashing (DSSH), which tackles the challenging problem from two perspectives. Firstly, we propose an intuitive 3D shape representation method to deal with unaligned shapes with arbitrary poses. Specifically, the proposed Segmented Stochastic-viewing Shape Network models discriminative 3D representations by a set of 2D images rendered from multiple views, which are stochastically selected from non-overlapping spatial segments of a 3D sphere. Secondly, Batch-Hard Binary Coding (BHBC) is developed to learn semantics-preserving compact binary codes by mining the hardest samples. The overall framework is jointly learned by developing an alternating iteration algorithm. Extensive experimental results on three benchmarks show that DSSH improves both the retrieval efficiency and accuracy remarkably, compared to the state-of-the-art methods.

Common object counting in a natural scene is a challenging problem in computer vision with numerous real-world applications. Existing image-level supervised common object counting approaches only predict the global object count and rely on additional instance-level supervision to also determine object locations. We propose an image-level supervised approach that provides both the global object count and the spatial distribution of object instances by constructing an object category density map. Motivated by psychological studies, we further reduce image-level supervision using a limited object count information (up to four). To the best of our knowledge, we are the first to propose image-level supervised density map estimation for common object counting and demonstrate its effectiveness in image-level supervised instance segmentation. Comprehensive experiments are performed on the PASCAL VOC and COCO datasets. Our approach outperforms existing methods, including those using instance-level supervision, on both datasets for common object counting. Moreover, our approach improves state-of-the-art image-level supervised instance segmentation with a relative gain of 17.8% in terms of average best overlap, on the PASCAL VOC 2012 dataset.

Batch Normalization (BN) is ubiquitously employed for accelerating neural network training and improving the generalization capability by performing standardization within mini-batches. Decorrelated Batch Normalization (DBN) further boosts the above effectiveness by whitening. However, DBN relies heavily on either a large batch size, or eigen-decomposition that suffers from poor efficiency on GPUs. We propose Iterative Normalization (IterNorm), which employs Newton’s iterations for much more efficient whitening, while simultaneously avoiding the eigen-decomposition. Furthermore, we develop a comprehensive study to show IterNorm has better trade-off between optimization and generalization, with theoretical and experimental support. To this end, we exclusively introduce Stochastic Normalization Disturbance (SND), which measures the inherent stochastic uncertainty of samples when applied to normalization operations. With the support of SND, we provide natural explanations to several phenomena from the perspective of optimization, e.g., why group-wise whitening of DBN generally outperforms full-whitening and why the accuracy of BN degenerates with reduced batch sizes. We demonstrate the consistently improved performance of IterNorm with extensive experiments on CIFAR-10 and ImageNet over BN and DBN.

Abnormal event detection in video is a challenging vision problem. Most existing approaches formulate abnormal event detection as an outlier detection task, due to the scarcity of anomalous data during training. Because of the lack of prior information regarding abnormal events, these methods are not fully-equipped to differentiate between normal and abnormal events. In this work, we formalize abnormal event detection as a one-versus-rest binary classification problem. Our contribution is two-fold. First, we introduce an unsupervised feature learning framework based on object-centric convolutional auto-encoders to encode both motion and appearance information. Second, we propose a supervised classification approach based on clustering the training samples into normality clusters. A one-versus-rest abnormal event classifier is then employed to separate each normality cluster from the rest. For the purpose of training the classifier, the other clusters act as dummy anomalies. During inference, an object is labeled as abnormal if the highest classification score assigned by the one-versus-rest classifiers is negative. Comprehensive experiments are performed on four benchmarks: Avenue, ShanghaiTech, UCSD and UMN. Our approach provides superior results on all four data sets. On the large-scale ShanghaiTech data set, our method provides an absolute gain of 8.4% in terms of frame-level AUC compared to the state-of-the-art method.

Crowd counting has recently attracted increasing interest in computer vision but remains a challenging problem. In this paper, we propose a trellis encoder-decoder network (TEDnet) for crowd counting, which focuses on generating high-quality density estimation maps. The major contributions are four-fold. First, we develop a new trellis architecture that incorporates multiple decoding paths to hierarchically aggregate features at different encoding stages, which improves the representative capability of convolutional features for large variations in objects. Second, we employ dense skip connections interleaved across paths to facilitate sufficient multi-scale feature fusions, which also helps TEDnet to absorb the supervision information. Third, we propose a new combinatorial loss to enforce similarities in local coherence and spatial correlation between maps. By distributedly imposing this combinatorial loss on intermediate outputs, TEDnet can improve the back-propagation process and alleviate the gradient vanishing problem. Finally, on four widely-used benchmarks, our TEDnet achieves the best overall performance in terms of both density map quality and counting accuracy, with an improvement up to 14% in MAE metric. These results validate the effectiveness of TEDnet for crowd counting.

Learning unbiased models on imbalanced datasets is a significant challenge. Rare classes tend to get a concentrated representation in the classification space which hampers the generalization of learned boundaries to new test examples. In this paper, we demonstrate that the Bayesian uncertainty estimates directly correlate with the rarity of classes and the difficulty level of individual samples. Subsequently, we present a novel framework for uncertainty based class imbalance learning that follows two key insights: First, classification boundaries should be extended further away from a more uncertain (rare) class to avoid over-fitting and enhance its generalization. Second, each sample should be modeled as a multi-variate Gaussian distribution with a mean vector and a covariance matrix defined by the sample's uncertainty. The learned boundaries should respect not only the individual samples but also their distribution in the feature space. Our proposed approach efficiently utilizes sample and class uncertainty information to learn robust features and more generalizable classifiers. We systematically study the class imbalance problem and derive a novel loss formulation for max-margin learning based on Bayesian uncertainty measure. The proposed method shows significant performance improvements on six benchmark datasets for face verification, attribute prediction, digit/object classification and skin lesion detection.

We introduce a novel network, called as CO-attention Siamese Network (COSNet), to address the unsupervised video object segmentation task from a holistic view. We emphasize the importance of inherent correlation among video frames and incorporate a global co-attention mechanism to improve further the state-of-the-art deep learning based solutions that primarily focus on learning discriminative foreground representations over appearance and motion in short-term temporal segments. The co-attention layers in our network provide efficient and competent stages for capturing global correlations and scene context by jointly computing and appending co-attention responses into a joint feature space. We train COSNet with pairs of video frames, which naturally augments training data and allows increased learning capacity. During the segmentation stage, the co-attention model encodes useful information by processing multiple reference frames together, which is leveraged to infer the frequently reappearing and salient foreground objects better. We propose a unified and end-to-end trainable framework where different co-attention variants can be derived for mining the rich context within videos. Our extensive experiments over three large benchmarks manifest that COSNet outperforms the current alternatives by a large margin. We will publicly release our implementation and models.

Generalized zero-shot action recognition is a challenging problem, where the task is to recognize new action categories that are unavailable during the training stage, in addition to the seen action categories. Existing approaches suffer from the inherent bias of the learned classifier towards the seen action categories. As a consequence, unseen category samples are incorrectly classified as belonging to one of the seen action categories. In this paper, we set out to tackle this issue by arguing for a separate treatment of seen and unseen action categories in generalized zero-shot action recognition. We introduce an out-of-distribution detector that determines whether the video features belong to a seen or unseen action category. To train our out-of-distribution detector, video features for unseen action categories are synthesized using generative adversarial networks trained on seen action category features. To the best of our knowledge, we are the first to propose an out-of-distribution detector based GZSL framework for action recognition in videos. Experiments are performed on three action recognition datasets: Olympic Sports, HMDB51 and UCF101. For generalized zero-shot action recognition, our proposed approach outperforms the baseline with absolute gains (in classification accuracy) of 7.0%, 3.4%, and 4.9%, respectively, on these datasets.

Single-stage object detectors have recently gained popularity due to their combined advantage of high detection accuracy and real-time speed. However, while promising results have been achieved by these detectors on standard-sized objects, their performance on small objects is far from satisfactory. To detect very small/large objects, classical pyramid representation can be exploited, where an image pyramid is used to build a feature pyramid (featurized image pyramid), enabling detection across a range of scales. Existing single-stage detectors avoid such a featurized image pyramid representation due to its memory and time complexity. In this paper, we introduce a light-weight architecture to efficiently produce featurized image pyramid in a single-stage detection framework. The resulting multi-scale features are then injected into the prediction layers of the detector using an attention module. The performance of our detector is validated on two benchmarks: PASCAL VOC and MS COCO. For a 300×300 input, our detector operates at 111 frames per second (FPS) on a Titan X GPU, providing state-of-the-art detection accuracy on PASCAL VOC 2007 testset. On the MS COCO testset, our detector achieves state-of-the-art results surpassing all existing single-stage methods in the case of single-scale inference.

This paper presents a salient object detection method that integrates both top-down and bottom-up saliency inference in an iterative and cooperative manner. The top-down process is used for coarse-to-fine saliency estimation, where high-level saliency is gradually integrated with finer lower-layer features to obtain a fine-grained result. The bottom-up process infers the high-level, but rough saliency through gradually using upper-layer, semantically-richer features. These two processes are alternatively performed, where the bottom-up process uses the fine-grained saliency obtained from the top-down process to yield enhanced high-level saliency estimate, and the top-down process, in turn, is further benefited from the improved high-level information. The network layers in the bottom-up/top-down processes are equipped with recurrent mechanisms for layer-wise, step-by-step optimization. Thus, saliency information is effectively encouraged to flow in a bottom-up, top-down and intra-layer manner. We show that most other saliency models based on fully convolutional networks (FCNs) are essentially variants of our model. Extensive experiments on several famous benchmarks clearly demonstrate the superior performance, good generalization, and powerful learning ability of our proposed saliency inference framework.

Semantic segmentation is an important computer vision task, which aims to allocate a semantic label to each pixel in an image. When training a segmentation model, it is common to fine-tune a classification network pre-trained on a large-scale dataset. However, as an intrinsic property of the classification model, invariance to spatial perturbation resulting from the lose of detail-sensitivity prevents segmentation networks from achieving high performance. The use of standard poolings is one of the key factors for this invariance. The most common standard poolings are max and average pooling. Max pooling can increase both the invariance to spatial perturbations and the non-linearity of the networks. Average pooling, on the other hand, is sensitive to spatial perturbations, but is a linear function. For semantic segmentation, we prefer both the preservation of detailed cues within a local feature region and non-linearity that increases a network's functional complexity. In this work, we propose a polynomial pooling (P-pooling) function that finds an intermediate form between max and average pooling to provide an optimally balanced and self-adjusted pooling strategy for semantic segmentation. The P-pooling is differentiable and can be applied into a variety of pre-trained networks. Extensive studies on the PASCAL VOC, Cityscapes and ADE20k datasets demonstrate the superiority of P-pooling over other poolings. Experiments on various network architectures and state-of-the-art training strategies also show that models with P-pooling layers consistently outperform those directly fine-tuned using pre-trained classification models.

Zero-shot learning (ZSL) aims to classify images from unseen categories, by merely utilizing seen class images as the training data. Existing works on ZSL mainly leverage the global features or learn the global regions, from which, to construct the embeddings to the semantic space. However, few of them study the discrimination power implied in local image regions (parts), which, in some sense, correspond to semantic attributes, have stronger discrimination than attributes, and can thus assist the semantic transfer between seen/unseen classes. In this paper, to discover (semantic) regions, we propose the attentive region embedding network (AREN), which is tailored to advance the ZSL task. Specifically, AREN is end-to-end trainable and consists of two network branches, i.e., the attentive region embedding (ARE) stream, and the attentive compressed second-order embedding (ACSE) stream. ARE is capable of discovering multiple part regions under the guidance of the attention and the compatibility loss. Moreover, a novel adaptive thresholding mechanism is proposed for suppressing redundant (such as background) attention regions. To further guarantee more stable semantic transfer from the perspective of second-order collaboration, ACSE is incorporated into the AREN. In the comprehensive evaluations on four benchmarks, our models achieve state-of-the-art performances under ZSL setting, and compelling results under generalized ZSL setting.

Medical image analysis has two important research areas: disease grading and fine-grained lesion segmentation. Although the former problem often relies on the latter, the two are usually studied separately. Disease severity grading can be treated as a classification problem, which only requires image-level annotations, while the lesion segmentation requires stronger pixel-level annotations. However, pixel-wise data annotation for medical images is highly time-consuming and requires domain experts. In this paper, we propose a collaborative learning method to jointly improve the performance of disease grading and lesion segmentation by semi-supervised learning with an attention mechanism. Given a small set of pixel-level annotated data, a multi-lesion mask generation model first performs the traditional semantic segmentation task. Then, based on initially predicted lesion maps for large quantities of image-level annotated data, a lesion attentive disease grading model is designed to improve the severity classification accuracy. Meanwhile, the lesion attention model can refine the lesion maps using class-specific information to fine-tune the segmentation model in a semi-supervised manner. An adversarial architecture is also integrated for training. With extensive experiments on a representative medical problem called diabetic retinopathy (DR), we validate the effectiveness of our method and achieve consistent improvements over state-of-the-art methods on three public datasets.

These years have seen the advance of compressive sensing (CS), but the CS-based image coding scheme still has a poor rate-distortion (R-D) performance compared with the traditional image coding techniques. In this paper, we propose an image coding scheme based on the CS paradigm via multi-layer closed-loop prediction. In the scheme, we divide CS measurements into multi-layers and predict a particular layer's measurements with all its preceding layers' measurements, which can reduce the redundancies between CS measurements efficiently. The produced measurement residuals are then quantized into binary codes, which are tremendously reduced compared to quantizing the CS measurements directly. Furthermore, We provide a non-local low-rank CS reconstruction algorithm corresponding to our multi-layer closed-loop prediction scheme. Experimental results verify that the proposed scheme can significantly outperform JPEG2000, and the reconstruction quality of our scheme is no worse or even better than that of HEVC-Intra.

The outcome of standard statistical shape modelling is a vector space representation of objects. Any convex combination of vectors of a set of object class examples generates a real and valid example. In this paper, we propose a Point Auto-Encoder (PointAE) with skip-connection, attention blocks for 3D statistical shape modelling directly on 3D points. The proposed PointAE is able to refine the correspondence with a correspondence refinement block. The data with refined correspondence can be fed to the PointAE again and bootstrap the constructed statistical models. Instead of two seperate models, PointAE can simultaneously model the shape and texture variation. The extensive evaluation in three open-sourced datasets demonstrates that the proposed method achieves better performance in representation ability of the shape variations.

Real-world object classes appear in imbalanced ratios. This poses a significant challenge for classifiers which get biased towards frequent classes. We hypothesize that improving the generalization capability of a classifier should improve learning on imbalanced datasets. Here, we introduce the first hybrid loss function that jointly performs classification and clustering in a single formulation. Our approach is based on an 'affinity measure' in Euclidean space that leads to the following benefits: (1) direct enforcement of maximum margin constraints on classification boundaries, (2) a tractable way to ensure uniformly spaced and equidistant cluster centers, (3) flexibility to learn multiple class prototypes to support diversity and discriminability in feature space. Our extensive experiments demonstrate the significant performance improvements on visual classification and verification tasks on multiple imbalanced datasets. The proposed loss can easily be plugged in any deep architecture as a differentiable block and demonstrates robustness against different levels of data imbalance and corrupted labels.

Deep neural networks are vulnerable to adversarial attacks which can fool them by adding minuscule perturbations to the input images. The robustness of existing defenses suffers greatly under white-box attack settings, where an adversary has full knowledge about the network and can iterate several times to find strong perturbations. We observe that the main reason for the existence of such perturbations is the close proximity of different class samples in the learned feature space. This allows model decisions to be totally changed by adding an imperceptible perturbation in the inputs. To counter this, we propose to class-wise disentangle the intermediate feature representations of deep networks. Specifically, we force the features for each class to lie inside a convex polytope that is maximally separated from the polytopes of other classes. In this manner, the network is forced to learn distinct and distant decision regions for each class. We observe that this simple constraint on the features greatly enhances the robustness of learned models, even against the strongest white-box attacks, without degrading the classification performance on clean images. We report extensive evaluations in both black-box and white-box attack scenarios and show significant gains in comparison to state-of-the art defenses.

Temporal action localization is a challenging computer vision problem with numerous real-world applications. Most existing methods require laborious frame-level supervision to train action localization models. In this work, we propose a framework, called 3C-Net, which only requires video-level supervision (weak supervision) in the form of action category labels and the corresponding count. We introduce a novel formulation to learn discriminative action features with enhanced localization capabilities. Our joint formulation has three terms: a classification term to ensure the separability of learned action features, an adapted multi-label center loss term to enhance the action feature discriminability and a counting loss term to delineate adjacent action sequences, leading to improved localization. Comprehensive experiments are performed on two challenging benchmarks: THUMOS14 and ActivityNet 1.2. Our approach sets a new state-of-the-art for weakly-supervised temporal action localization on both datasets. On the THUMOS14 dataset, the proposed method achieves an absolute gain of 4.6% in terms of mean average precision (mAP), compared to the state-of-the-art. Source code is available at https://github.com/naraysa/3c-net.

We propose a single-stage detection framework that jointly tackles the problem of multi-scale object detection and class imbalance. Rather than designing deeper networks, we introduce a simple yet effective feature enrichment scheme to produce multi-scale contextual features. We further introduce a cascaded refinement scheme which first instills multi-scale contextual features into the prediction layers of the single-stage detector in order to enrich their discriminative power for multi-scale detection. Second, the cascaded refinement scheme counters the class imbalance problem by refining the anchors and enriched features to improve classification and regression. Experiments are performed on two benchmarks: PASCAL VOC and MS COCO. For a 320×320 input on the MS COCO test-dev, our detector achieves state-of-the-art single-stage detection accuracy with a COCO AP of 33.2 in the case of single-scale inference, while operating at 21 milliseconds on a Titan XP GPU. For a 512×512 input on the MS COCO test-dev, our approach obtains an absolute gain of 1.6% in terms of COCO AP, compared to the best reported single-stage results[5]. Source code and models are available at: https://github.com/Ranchentx/EFGRNet.

Aggregating multi-level features is essential for capturing multi-scale context information for precise scene semantic segmentation. However, the improvement by directly fusing shallow features and deep features becomes limited as the semantic gap between them increases. To solve this problem, we explore two strategies for robust feature fusion. One is enhancing shallow features using a semantic enhancement module (SeEM) to alleviate the semantic gap between shallow features and deep features. The other strategy is feature attention, which involves discovering complementary information (i.e., boundary information) from low-level features to enhance high-level features for precise segmentation. By embedding these two strategies, we construct a parallel feature pyramid towards improving multi-level feature fusion. A Semantic Enhanced Network called SeENet is constructed with the parallel pyramid to implement precise segmentation. Experiments on three benchmark datasets demonstrate the effectiveness of our method for robust multi-level feature aggregation. As a result, our SeENet has achieved better performance than other state-of-the-art methods for semantic segmentation.

Pedestrian detection relying on deep convolution neural networks has made significant progress. Though promising results have been achieved on standard pedestrians, the performance on heavily occluded pedestrians remains far from satisfactory. The main culprits are intra-class occlusions involving other pedestrians and inter-class occlusions caused by other objects, such as cars and bicycles. These results in a multitude of occlusion patterns. We propose an approach for occluded pedestrian detection with the following contributions. First, we introduce a novel mask-guided attention network that fits naturally into popular pedestrian detection pipelines. Our attention network emphasizes on visible pedestrian regions while suppressing the occluded ones by modulating full body features. Second, we empirically demonstrate that coarse-level segmentation annotations provide reasonable approximation to their dense pixel-wise counterparts. Experiments are performed on CityPersons and Caltech datasets. Our approach sets a new state-of-the-art on both datasets. Our approach obtains an absolute gain of 9.5% in log-average miss rate, compared to the best reported results [32] on the heavily occluded HO pedestrian set of CityPersons test set. Further, on the HO pedestrian set of Caltech dataset, our method achieves an absolute gain of 5.0% in log-average miss rate, compared to the best reported results [13]. Code and models are available at: https://github.com/Leotju/MGAN.

This paper proposes a human-aware deblurring model that disentangles the motion blur between foreground (FG) humans and background (BG). The proposed model is based on a triple-branch encoder-decoder architecture. The first two branches are learned for sharpening FG humans and BG details, respectively; while the third one produces global, harmonious results by comprehensively fusing multi-scale deblurring information from the two domains. The proposed model is further endowed with a supervised, human-aware attention mechanism in an end-to-end fashion. It learns a soft mask that encodes FG human information and explicitly drives the FG/BG decoder-branches to focus on their specific domains. Above designs lead to a fully differentiable motion deblurring network, which can be trained end-to-end. To further benefit the research towards Human-aware Image Deblurring, we introduce a large-scale dataset, named HIDE, which consists of 8,422 blurry and sharp image pairs with 65,784 densely annotated FG human bounding boxes. HIDE is specifically built to span a broad range of scenes, human object sizes, motion patterns, and background complexities. Extensive experiments on public benchmarks and our dataset demonstrate that our model performs favorably against the state-of-the-art motion deblurring methods, especially in capturing semantic details.

Single-stage object detection methods have received significant attention recently due to their characteristic realtime capabilities and high detection accuracies. Generally, most existing single-stage detectors follow two common practices: they employ a network backbone that is pretrained on ImageNet for the classification task and use a top-down feature pyramid representation for handling scale variations. Contrary to common pre-training strategy, recent works have demonstrated the benefits of training from scratch to reduce the task gap between classification and localization, especially at high overlap thresholds. However, detection models trained from scratch require significantly longer training time compared to their typical finetuning based counterparts. We introduce a single-stage detection framework that combines the advantages of both fine-tuning pretrained models and training from scratch. Our framework constitutes a standard network that uses a pre-trained backbone and a parallel light-weight auxiliary network trained from scratch. Further, we argue that the commonly used top-down pyramid representation only focuses on passing high-level semantics from the top layers to bottom layers. We introduce a bi-directional network that efficiently circulates both low-/mid-level and high-level semantic information in the detection framework. Experiments are performed on MS COCO and UAVDT datasets. Compared to the baseline, our detector achieives an absolute gain of 7.4% and 4.2% in average precision (AP) on MS COCO and UAVDT datasets, respectively using VGG backbone. For a 300×300 input on the MS COCO test set, our detector with ResNet backbone surpasses existing single-stage detection methods for single-scale inference achieving 34.3 AP, while operating at an inference time of 19 milliseconds on a single Titan X GPU. Code is avail- able at https://github.com/vaesl/LRF-Net.

Human-object interaction detection is an important and relatively new class of visual relationship detection tasks, essential for deeper scene understanding. Most existing approaches decompose the problem into object localization and interaction recognition. Despite showing progress, these approaches only rely on the appearances of humans and objects and overlook the available context information, crucial for capturing subtle interactions between them. We propose a contextual attention framework for human-object interaction detection. Our approach leverages context by learning contextually-aware appearance features for human and object instances. The proposed attention module then adaptively selects relevant instance-centric context information to highlight image regions likely to contain human-object interactions. Experiments are performed on three benchmarks: V-COCO, HICO-DET and HCVRD. Our approach outperforms the state-of-the-art on all datasets. On the V-COCO dataset, our method achieves a relative gain of 4.4% in terms of role mean average precision (mAP role ), compared to the existing best approach.

This work proposes a novel attentive graph neural network (AGNN) for zero-shot video object segmentation (ZVOS). The suggested AGNN recasts this task as a process of iterative information fusion over video graphs. Specifically, AGNN builds a fully connected graph to efficiently represent frames as nodes, and relations between arbitrary frame pairs as edges. The underlying pair-wise relations are described by a differentiable attention mechanism. Through parametric message passing, AGNN is able to efficiently capture and mine much richer and higher-order relations between video frames, thus enabling a more complete understanding of video content and more accurate foreground estimation. Experimental results on three video segmentation datasets show that AGNN sets a new state-of-the-art in each case. To further demonstrate the generalizability of our framework, we extend AGNN to an additional task: image object co-segmentation (IOCS). We perform experiments on two famous IOCS datasets and observe again the superiority of our AGNN model. The extensive experiments verify that AGNN is able to learn the underlying semantic/appearance relationships among video frames or related images, and discover the common objects.

Despite online learning (OL) techniques have boosted the performance of semi-supervised video object segmentation (VOS) methods, the huge time costs of OL greatly restricts their practicality. Matching based and propagation based methods run at a faster speed by avoiding OL techniques. However, they are limited by sub-optimal accuracy, due to mismatching and drifting problems. In this paper, we develop a real-time yet very accurate Ranking Attention Network (RANet) for VOS. Specifically, to integrate the insights of matching based and propagation based methods, we employ an encoder-decoder framework to learn pixel-level similarity and segmentation in an end-to-end manner. To better utilize the similarity maps, we propose a novel ranking attention module, which automatically ranks and selects these maps for fine-grained VOS performance. Experiments on DAVIS16 and DAVIS17 datasets show that our RANet achieves the best speed-accuracy trade-off, e.g., with 33 milliseconds per frame and J&F=85.5% on DAVIS16. With OL, our RANet reaches J&F=87.1% on DAVIS16, exceeding state-of-the-art VOS methods. The code can be found at https://github.com/Storife/RANet.

This work proposes to combine neural networks with the compositional hierarchy of human bodies for efficient and complete human parsing. We formulate the approach as a neural information fusion framework. Our model assembles the information from three inference processes over the hierarchy: direct inference (directly predicting each part of a human body using image information), bottom-up inference (assembling knowledge from constituent parts), and top-down inference (leveraging context from parent nodes). The bottom-up and top-down inferences explicitly model the compositional and decompositional relations in human bodies, respectively. In addition, the fusion of multi-source information is conditioned on the inputs, i.e., by estimating and considering the confidence of the sources. The whole model is end-to-end differentiable, explicitly modeling information flows and structures. Our approach is extensively evaluated on four popular datasets, outperforming the state-of-the-arts in all cases, with a fast processing speed of 23fps. Our code and results have been released to help ease future research in this direction.

Crowd counting is receiving rapidly growing research interests due to its potential application value in numerous real-world scenarios. However, due to various challenges such as occlusion, insufficient resolution and dynamic backgrounds, crowd counting remains an unsolved problem in computer vision. Density estimation is a popular strategy for crowd counting, where conventional density estimation methods perform pixel-wise regression without explicitly accounting the interdependence of pixels. As a result, independent pixel-wise predictions can be noisy and inconsistent. In order to address such an issue, we propose a Relational Attention Network (RANet) with a self-attention mechanism for capturing interdependence of pixels. The RANet enhances the self-attention mechanism by accounting both short-range and long-range interdependence of pixels, where we respectively denote these implementations as local self-attention (LSA) and global self-attention (GSA). We further introduce a relation module to fuse LSA and GSA to achieve more informative aggregated feature representations. We conduct extensive experiments on four public datasets, including ShanghaiTech A, ShanghaiTech B, UCF-CC-50 and UCF-QNRF. Experimental results on all datasets suggest RANet consistently reduces estimation errors and surpasses the state-of-the-art approaches by large margins.

Crowd counting has recently generated huge popularity in computer vision, and is extremely challenging due to the huge scale variations of objects. In this paper, we propose the Attentional Neural Field (ANF) for crowd counting via density estimation. Within the encoder-decoder network, we introduce conditional random fields (CRFs) to aggregate multi-scale features, which can build more informative representations. To better model pair-wise potentials in CRFs, we incorperate non-local attention mechanism implemented as inter- and intra-layer attentions to expand the receptive field to the entire image respectively within the same layer and across different layers, which captures long-range dependencies to conquer huge scale variations. The CRFs coupled with the attention mechanism are seamlessly integrated into the encoder-decoder network, establishing an ANF that can be optimized end-to-end by back propagation. We conduct extensive experiments on four public datasets, including ShanghaiTech, WorldEXPO 10, UCF-CC-50 and UCF-QNRF. The results show that our ANF achieves high counting performance, surpassing most previous methods.

A set containment join operates on two set-valued attributes with a subset (⊆) relationship as the join condition. It has many real-world applications, such as in publish/subscribe services and inclusion dependency discovery. Existing solutions can be broadly classified into union-oriented and intersection-oriented methods. Based on several recent studies, union-oriented methods are not competitive as they involve an expensive subset enumeration step. Intersection-oriented methods build an inverted index on one attribute and perform inverted list intersection on another attribute. Existing intersection-oriented methods intersect inverted lists one-by-one. In contrast, in this paper, we propose to intersect all the inverted lists simultaneously while skipping many irrelevant entries in the lists. To share computation, we utilize the prefix tree structure and extend our novel list intersection method to operate on the prefix tree. To further improve the efficiency, we propose to partition the data and use different methods to process each partition. We evaluated our methods using both real-world and synthetic datasets. Experimental results show that our approach outperforms existing methods by up to 10×.

In this paper, we investigate the structural similarities within a finite Markov decision process (MDP). We view a finite MDP as a heterogeneous directed bipartite graph and propose novel measures for state similarity and action similarity in a mutual reinforcement manner. We prove that the state similarity is a metric and the action similarity is a pseudometric. We also establish the connection between the proposed similarity measures and the optimal values of the MDP. Extensive experiments show that the proposed measures are effective.

With the popularization of portable devices, numerous applications continuously produce huge streams of geo-tagged textual data, thus posing challenges to index geo-textual streaming data efficiently, which is an important task in both data management and AI applications, e.g., real-time data streams mining and targeted advertising. This, however, is not possible with the state-of-the-art indexing methods as they focus on search optimizations of static datasets, and have high index maintenance cost. In this paper, we present NQ-tree, which combines new structure designs and self-tuning methods to navigate between update and search efficiency. Our contributions include: (1) the design of multiple stores each with a different emphasis on write-friendness and read-friendness; (2) utilizing data compression techniques to reduce the I/O cost; (3) exploiting both spatial and keyword information to improve the pruning efficiency; (4) proposing an analytical cost model, and using an online self-tuning method to achieve efficient accesses to different workloads. Experiments on two real-world datasets show that NQ-tree outperforms two well designed baselines by up to 10×.

Recently, deep learning-based single image dehazing method has been a popular approach to tackle dehazing. However, the existing dehazing approaches are performed directly on the original hazy image, which easily results in image blurring and noise amplifying. To address this issue, the paper proposes a DPDP-Net (Dual-Path in Dual-Path network) framework by employing a hierarchical dual path network. Specifically, the first-level dual-path network consists of a Dehazing Network and a Denoising Network, where the Dehazing Network is responsible for haze removal in the structural layer, and the Denoising Network deals with noise in the textural layer, respectively. And the second-level dual-path network lies in the Dehazing Network, which has an AL-Net (Atmospheric Light Network) and a TM-Net (Transmission Map Network), respectively. Concretely, the AL-Net aims to train the non-uniform atmospheric light, while the TM-Net aims to train the transmission map that reflects the visibility of the image. The final dehazing image is obtained by nonlinearly fusing the output of the Denoising Network and the Dehazing Network. Extensive experiments demonstrate that our proposed DPDP-Net achieves competitive performance against the state-of-the-art methods on both synthetic and real-world images.

Due to the high cost of manual annotation, learning directly from the web has attracted broad attention. One issue that limits their performance is the problem of visual polysemy. To address this issue, we present an adaptive multi-model framework that resolves polysemy by visual disambiguation. Compared to existing methods, the primary advantage of our approach lies in that our approach can adapt to the dynamic changes in the search results. Our proposed framework consists of two major steps: we first discover and dynamically select the text queries according to the image search results, then we employ the proposed saliency-guided deep multi-instance learning network to remove outliers and learn classification models for visual disambiguation. Extensive experiments demonstrate the superiority of our proposed approach.

With the increasing availability of moving-object tracking data, use of this data for route search and recommendation is increasingly important. To this end, we propose a novel parallel split-and-combine approach to enable route search by locations (RSL-Psc). Given a set of routes, a set of places to visit O, and a threshold θ, we retrieve the route composed of sub-routes that (i) has similarity to O no less than θ and (ii) contains the minimum number of sub-route combinations. The resulting functionality targets a broad range of applications, including route planning and recommendation, ridesharing, and location-based services in general. To enable efficient and effective RSL-Psc computation on massive route data, we develop novel search space pruning techniques and enable use of the parallel processing capabilities of modern processors. Specifically, we develop two parallel algorithms, Fully-Split Parallel Search (FSPS) and Group-Split Parallel Search (GSPS). We divide the route split-and-combine task into ∑k=0 M S(|O|,k+1) sub-tasks, where M is the maximum number of combinations and S(⋅) is the Stirling number of the second kind. In each sub-task, we use network expansion and exploit spatial similarity bounds for pruning. The algorithms split candidate routes into sub-routes and combine them to construct new routes. The sub-tasks are independent and are performed in parallel. Extensive experiments with real data offer insight into the performance of the algorithms, indicating that our RSL-Psc problem can generate high-quality results and that the two algorithms are capable of achieving high efficiency and scalability.

In the literature of video analysis, most researches, such as retrieval and recognition, hypothesize that each input video contains at least one complete semantic entity, e.g. an activity, action and event.However, this hypothesis does not hold in many realistic scenarios due to two main reasons. First, complete videos whose qualities are good enough for automatic analysis are not always accessible because of heavy motion blur, occlusions, interruptions, etc. % Second, extracting features from complete videos always fails to meet up with speed and storage requirements in large-scale use cases.To tackle these challenges, incomplete videos are more useful, but researches on them are seldom mentioned. In this paper, we propose a novel and effective hashing framework specialized in large-scale incomplete video analysis called Generative Reconstructive Hashing (GRH). To begin with, an adversarial generative network that is specially designed to map incomplete video features to the feature distributions of complete videos, so that features of incomplete videos become indistinguishable from those of complete videos. Then, the discriminative hashing module further fills the gap between full video features and estimated features from partial videos by projecting both features into a common binary feature space, which allows improvement in efficiency compared with real-value based methods. GRH is the first end-to-end framework for incomplete video analysis. Extensive experiments on various datasets demonstrate GRH's superior effectiveness and efficiency on retrieval and recognition tasks. GRH outperforms the recent state-of-the-art methods by 5.44/3.22/4.82 in terms of MAPs on HMDB51/UCF101/CCV datasets, respectively.

Learning the probability distribution of high-dimensional data is a challenging problem. To solve this problem, we formulate a deep energy adversarial network (DEAN), which casts the energy model learned from real data into an optimization of a goodness-of-fit (GOF) test statistic. DEAN can be interpreted as a GOF game between two generative networks, where one explicit generative network learns an energy-based distribution that fits the real data, and the other implicit generative network is trained by minimizing a GOF test statistic between the energy-based distribution and the generated data, such that the underlying distribution of the generated data is close to the energy-based distribution. We design a two-level alternative optimization procedure to train the explicit and implicit generative networks, such that the hyper-parameters can also be automatically learned. Experimental results show that DEAN achieves high quality generations compared to the state-of-the-art approaches.

Incremental life-long learning is a main challenge towards the long-standing goal of Artificial General Intelligence. In real-life settings, learning tasks arrive in a sequence and machine learning models must continually learn to increment already acquired knowledge. The existing incremental learning approaches fall well below the state-of-the-art cumulative models that use all training classes at once. In this paper, we propose a random path selection algorithm, called RPS-Net, that progressively chooses optimal paths for the new tasks while encouraging parameter sharing and reuse. Our approach avoids the overhead introduced by computationally expensive evolutionary and reinforcement learning based path selection strategies while achieving considerable performance gains. As an added novelty, the proposed model integrates knowledge distillation and retrospection along with the path selection strategy to overcome catastrophic forgetting. In order to maintain an equilibrium between previous and newly acquired knowledge, we propose a simple controller to dynamically balance the model plasticity. Through extensive experiments, we demonstrate that the proposed method surpasses the state-of-the-art performance on incremental learning and by utilizing parallel computation this method can run in constant time with nearly the same efficiency as a conventional deep convolutional neural network.

Makeup face verification in the wild is an important research problem for its popularization in real-world. However, little effort has been made to tackle it in computer vision. In this research, we first build a new database, i.e., Facial Beauty Database (FBD), which contains paired facial images of 8933 subjects without and with makeup in different real-world scenarios. To the best of our knowledge, FBD is the largest makeup face database to date compared with existing databases for facial makeup research. Moreover, we propose a new discriminative marginal metric learning (DMML) algorithm to deal with this problem in the wild. Inspired by the fact that interclass marginal faces are usually more discriminative than interclass nonmarginal faces in learning the discriminative metric space, we use the interclass marginal faces to depict the discriminative information. Simultaneously, we wish that those interclass marginal faces without makeup relations are separated from each other as far as possible, so that more discriminative information between facial images without and with makeup can be exploited for verification. Furthermore, since multiple features could provide comprehensive information in describing the facial representations from diverse points of view and extract more informative cues from facial images, we also introduce a multiview discriminative marginal metric learning (MDMML) algorithm by effectively learning a robust metric space such that multiple features from different points of view can be integrated to effectively enhance the performance of makeup face verification. Experimental results on two real-world makeup face databases are utilized to show the effectiveness of our method and the possibility of verifying the makeup relations from facial images in real-world.

Clustering is a long-standing important research problem, however, remains challenging when handling large-scale image data from diverse sources. In this paper, we present a novel Binary Multi-View Clustering (BMVC) framework, which can dexterously manipulate multi-view image data and easily scale to large data. To achieve this goal, we formulate BMVC by two key components: compact collaborative discrete representation learning and binary clustering structure learning, in a joint learning framework. Specifically, BMVC collaboratively encodes the multi-view image descriptors into a compact common binary code space by considering their complementary information; the collaborative binary representations are meanwhile clustered by a binary matrix factorization model, such that the cluster structures are optimized in the Hamming space by pure, extremely fast bit-operations. For efficiency, the code balance constraints are imposed on both binary data representations and cluster centroids. Finally, the resulting optimization problem is solved by an alternating optimization scheme with guaranteed fast convergence. Extensive experiments on four large-scale multi-view image datasets demonstrate that the proposed method enjoys the significant reduction in both computation and memory footprint, while observing superior (in most cases) or very competitive performance, in comparison with state-of-the-art clustering methods.

Hashing algorithm has been widely used to speed up image retrieval due to its compact binary code and fast distance calculation. The combination with deep learning boosts the performance of hashing by learning accurate representations and complicated hashing functions. So far, the most striking success in deep hashing have mostly involved discriminative models, which require labels. To apply deep hashing on datasets without labels, we propose a deep self-taught hashing algorithm (DSTH), which generates a set of pseudo labels by analyzing the data itself, and then learns the hash functions for novel data using discriminative deep models. Furthermore, we generalize DSTH to support both supervised and unsupervised cases by adaptively incorporating label information. We use two different deep learning framework to train the hash functions to deal with out-of-sample problem and reduce the time complexity without loss of accuracy. We have conducted extensive experiments to investigate different settings of DSTH, and compared it with state-of-the-art counterparts in six publicly available datasets. The experimental results show that DSTH outperforms the others in all datasets.

Currently, the most widely used point trajectories generation methods estimate the trajectories from the dense optical flow, by using a consistency check strategy to detect the occluded regions. However, these methods will miss some important trajectories, thus resulting in breaking smooth areas without any structure especially around the motion boundaries (MBs). We suggest exploring MBs in video to generate more accurate dense point trajectories. Estimating MBs from the video improves the point trajectory accuracy of the discontinuity or occluded areas. Then, we obtain trajectories by tracking the initial feature points through all frames. The experimental results demonstrate that our method outperforms the state-of-the-art methods on the challenging benchmark.

We present an unsupervised segmentation framework for stereo videos using stereoscopic trajectories. The proposed stereo trajectory shows favorable properties for modeling the long-term motion information through the whole sequence and explicitly capturing the corresponding relationships between two stereo views. The stereo prior is important for inferring the desired object and guarantees the consistent spatial-temporal segmentation, which contributes to an enjoyable stereo experience. We start by deriving stereo trajectories from left and right views simultaneously, which are represented via a graph structure. Then we detect object-like stereo trajectories via the graph structure to efficiently infer the desired object. Finally, an energy optimization function is proposed to produce the stereo segmentation results via leveraging the object information from stereo trajectories. To benefit potential research, we collected a new stereoscopic video benchmark, which consists of a total of 50 stereo video clips and includes many challenges in segmentation. Extensive experimental results demonstrate that our stereo segmentation method achieves higher performance and preserves better stereo structures, compared with prevailing competitors. The source code and results are available at: https://github.com/shenjianbing/StereoSeg.

In this paper, we propose a new multiobject visual tracking algorithm by submodular optimization. The proposed algorithm is composed of two main stages. At the first stage, a new selecting strategy of tracklets is proposed to cope with occlusion problem. We generate low-level tracklets using overlap criteria and min-cost flow, respectively, and then integrate them into a candidate tracklets set. In the second stage, we formulate the multiobject tracking problem as the submodular maximization problem subject to related constraints. The submodular function selects the correct tracklets from the candidate set of tracklets to form the object trajectory. Then, we design a connecting process which connects the corresponding trajectories to overcome the occlusion problem. Experimental results demonstrate the effectiveness of our tracking algorithm. Our source code is available at https://github.com/shenjianbing/submodulartrack.