Video diffusion models have achieved impressive results in natural scene generation, yet they struggle to generalize to scientific phenomena such as fluid simulations and meteorological processes, where underlying dynamics are governed by scientific laws. These tasks pose unique challenges, including severe domain gaps, limited training data, and the lack of descriptive language annotations. To handle this dilemma, we extracted the latent scientific phenomena knowledge and further proposed a fresh framework that teaches video diffusion models to generate scientific phenomena from a single initial frame. Particularly, static knowledge is extracted via pre-trained masked autoencoders, while dynamic knowledge is derived from pre-trained optical flow prediction. Subsequently, based on the aligned spatial relations between the CLIP vision and language encoders, the visual embeddings of scientific phenomena, guided by latent scientific phenomena knowledge, are projected to generate the pseudo-language prompt embeddings in both spatial and frequency domains. By incorporating these prompts and fine-tuning the video diffusion model, we enable the generation of videos that better adhere to scientific laws. Extensive experiments on both computational fluid dynamics simulations and real-world typhoon observations demonstrate the effectiveness of our approach, achieving superior fidelity and consistency across diverse scientific scenarios.

Vision foundation models (VFMs) have demonstrated remarkable capabilities in learning universal visual representations. However, adapting these models to downstream tasks conventionally requires parameter updates, with even parameter-efficient fine-tuning methods necessitating the modification of thousands to millions of weights. In this paper, we investigate the redundancies in the segment anything model (SAM) and then propose a novel parameter-free fine-tuning method. Unlike traditional fine-tuning methods that adjust parameters, our method emphasizes selecting, reusing, and enhancing pre-trained features, offering a new perspective on fine-tuning foundation models. Specifically, we introduce a channel selection algorithm based on the model's output difference to identify redundant and effective channels. By selectively replacing the redundant channels with more effective ones, we filter out less useful features and reuse more task-irrelevant features to downstream tasks, thereby enhancing the task-specific feature representation. Experiments on both out-of-domain and in-domain datasets demonstrate the efficiency and effectiveness of our method in different vision tasks (e.g., image segmentation, depth estimation and image classification). Notably, our approach can seamlessly integrate with existing fine-tuning strategies (e.g., LoRA, Adapter), further boosting the performance of already fine-tuned models. Moreover, since our channel selection involves only model inference, our method significantly reduces GPU memory overhead.

Score Distillation Sampling and its variants have shown strong potential in text-to-3D generation by leveraging scores estimated from pretrained text-to-image diffusion models to optimize 3D representations. However, due to the view-agnostic nature of these scores, existing methods often suffer from the multi-face Janus problem, leading to inconsistencies across different views. In this work, we propose Rectified Score Distillation, which addresses this issue by incorporating view-conditioned scores as priors. Specifically, we formulate a reverse ordinary differential equation (ODE) that is additionally conditioned on camera poses. Then we rectify the standard, view-irrelevant scores to approximate the desired gradients along this ODE. Building on these rectified scores, our full framework, named AgonicDreamer, enables the generation of photorealistic and multi-view consistent 3D content with fine-grained details, as validated by extensive experimental results.

Existing methods for learning 3D point cloud representation often use a single dataset-specific training and testing approach, leading to performance drops due to significant domain shifts between training and testing data. While recent cross-domain methods have made promising progress, the lack of inherent semantic information in point clouds makes models prone to overfitting specific datasets. As such, we introduce 3D-CFA, a simple yet effective cross-modality feature aggregation method for cross-domain 3D point cloud representation learning. 3D-CFA aggregates the geometry tokens with semantic tokens derived from multi-view images, which are projected from the point cloud, thus generating more transferable features for cross-domain 3D point cloud representation learning. Specifically, 3D-CFA consists of two main components: a cross-modality feature aggregation module and an elastic domain alignment module. The cross-modality feature aggregation module converts unordered points into multi-view images using the modality transformation module. Then, the geometry tokens and semantic tokens extracted from the geometry encoder and semantic encoder are fed into the cross-modal projector to get the transferable 3D tokens. A key insight of this design is that the semantic tokens can serve as a bridge between the 3D point cloud model and the 2D foundation model, greatly promoting the generalization of cross-domain models facing the severe domain shift. Finally, the elastic domain alignment module learns the hierarchical domain-invariant features of different training domains for either domain adaptation or domain generalization protocols. 3D-CFA finds a better way to transfer the knowledge of the 2D foundation model pre-trained at scale, meanwhile only introducing a few extra trainable parameters. Comprehensive experiments on several cross-domain point cloud benchmarks demonstrate the effectiveness of the proposed method compared to the state-of-the-art methods.

The advancement of generative artificial intelligence has led to the creation of more diverse and realistic fake facial images. This poses serious threats to personal privacy and can contribute to the spread of misinformation. Existing deepfake detection methods usually utilize prior knowledge about forged clues to design complex modules, achieving excellent performance in the intra-domain settings. However, their performance usually suffers from a significant decline in unseen forgery patterns. It is thus desirable to develop a generalized deepfake detection method using a neat network structure. In this paper, we propose a simple yet efficient framework to transfer a powerful large-scale vision model like ViT to the downstream deepfake detection task, namely the generalized deepfake detection framework (GenDF). Concretely, we first propose a deepfake-specific representation learning (DSRL) scheme to learn different discontinuity patterns across patches inside a fake facial image and continuity between patches within a real counterpart in a low-dimensional space. To further alleviate the distribution mismatch between generic real images and human facial images consisting of both real and fake, we introduce a feature space redistribution (FSR) scheme to separately optimize the distributions of real and fake feature space, enabling the model to learn more distinctive representations. Furthermore, to enhance the generalization performance on unseen forgery patterns produced by constantly evolving facial manipulation techniques and diverse variations on real faces, we propose a classification-invariant feature augmentation (CIFAug) function without trainable parameters. CIFAug expands the scopes of real and fake feature space along directions orthogonal to the classification direction, enabling the model to learn more generalizable features while preserving discrimination. Extensive experiments demonstrate that our method achieves state-of-the-art generalization performance in cross-domain and cross-manipulation settings with only 0.28M trainable parameters.

The Segment Anything Model (SAM) is a widely used vision foundation model with diverse applications, including image segmentation, detection, and tracking. Given SAM's wide applications, understanding its robustness against adversarial attacks is crucial for real-world deployment. However, research on SAM's robustness is still in its early stages. Existing attacks often overlook the role of prompts in evaluating SAM's robustness, and there has been insufficient exploration of defense methods to balance the robustness and accuracy. To address these gaps, this paper proposes an adversarial robustness framework designed to evaluate and enhance the robustness of SAM. Specifically, we introduce a cross-prompt attack method to enhance the attack transferability across different prompt types. Besides attacking, we propose a few-parameter adaptation strategy to defend SAM against various adversarial attacks. To balance robustness and accuracy, we use the singular value decomposition (SVD) to constrain the space of trainable parameters, where only singular values are adaptable. Experiments demonstrate that our cross-prompt attack method outperforms previous approaches in terms of attack success rate on both SAM and SAM 2. By adapting only 512 parameters, we achieve at least a 15% improvement in mean intersection over union (mIoU) against various adversarial attacks. Compared to previous defense methods, our approach enhances the robustness of SAM while maximally maintaining its original performance.

Recent 3D face reconstruction methods have made remarkable advancements, yet achieving high-quality facial reflectance from monocular input remains challenging. Existing methods rely on the light-stage captured data to learn facial reflectance models. However, limited subject diversity in these datasets poses challenges in achieving good generalization and broad applicability. This motivates us to explore whether the extensive priors captured in recent generative diffusion models (e.g., Stable Diffusion) can enable more generalizable facial reflectance estimation as these models have been pre-trained on large-scale internet image collections containing rich visual patterns. In this paper, we introduce the use of Stable Diffusion as a prior for facial reflectance estimation, achieving robust results with minimal captured data for fine-tuning. We present S3-Face, a comprehensive framework capable of producing SSS-compliant skin reflectance from in-the-wild images. Our method adopts a two-stage training approach: in the first stage, DSN-Net is trained to predict diffuse albedo, specular albedo, and normal maps from in-the-wild images using a novel joint reflectance attention module. In the second stage, HM-Net is trained to generate hemoglobin and melanin maps based on the diffuse albedo predicted in the first stage, yielding SSS-compliant and detailed reflectance maps. Extensive experiments demonstrate that our method achieves strong generalization and produces high-fidelity, SSS-compliant facial reflectance estimation.

Despite the advent in 3D hand pose estimation, current methods predominantly focus on single-image 3D hand reconstruction in the camera frame, overlooking the world-space motion of the hands. Such limitation prohibits their direct use in egocentric video settings, where hands and camera are continuously in motion. In this work, we propose HaWoR, a high-fidelity method for hand motion reconstruction in world coordinates from egocentric videos. We propose to decouple the task by reconstructing the hand motion in the camera space and estimating the camera trajectory in the world coordinate system. To achieve precise camera trajectory estimation, we propose an adaptive egocentric SLAM framework that addresses the shortcomings of traditional SLAM methods, providing robust performance under challenging camera dynamics. To ensure robust hand motion trajectories, even when the hands move out of view frustum, we devise a novel motion infiller net work that effectively completes the missing frames of the sequence. Through extensive quantitative and qualitative evaluations, we demonstrate that HaWoR achieves state-of-the-art performance on both hand motion reconstruction and world-frame camera trajectory estimation under different egocentric benchmark datasets. Code and models are available on https://hawor-project.github.io/.

Multimodal data is known to be helpful for visual tracking by improving robustness to appearance variations. However, sensor synchronization challenges often compromise data availability, particularly in video settings where shortages can be temporal. Despite its importance, this area remains underexplored. In this paper, we present the first comprehensive study on tracker performance with temporally incomplete multimodal data. Unsurprisingly, under such a circumstance, existing trackers exhibit significant performance degradation, as their rigid architectures lack the adaptability needed to effectively handle missing modalities. To address these limitations, we propose a flexible framework for robust multimodal tracking. We venture that a tracker should dynamically activate computational units based on missing data rates. This is achieved through a novel Heterogeneous Mixture-of-Experts fusion mechanism with adaptive complexity, coupled with a video-level masking strategy that ensures both temporal consistency and spatial completeness which is critical for effective video tracking. Surprisingly, our model not only adapts to varying missing rates but also adjusts to scene complexity. Extensive experiments show that our model achieves SOTA performance across 9 benchmarks, excelling in both conventional complete and missing modality settings. The code and benchmark will be publicly available at https://github.com/supertyd/FlexTrack/tree/main.

Recent progress in image generation has sparked research into controlling these models through condition signals, with various methods addressing specific challenges in conditional generation. Instead of proposing another specialized technique, we introduce a simple, unified framework to handle diverse conditional generation tasks involving a specific image-condition correlation. By learning a joint distribution over a correlated image pair (e.g. image and depth) with a diffusion model, our approach enables versatile capabilities via different inference-time sampling schemes, including controllable image generation (e.g. depth to image), estimation (e.g. image to depth), signal guidance, joint generation (image \& depth), and coarse control. Previous attempts at unification often introduce complexity through multi-stage training, architectural modification, or increased parameter counts. In contrast, our simplified formulation requires a single, computationally efficient training stage, maintains the standard model input, and adds minimal learned parameters (15% of the base model). Moreover, our model supports additional capabilities like non-spatially aligned and coarse conditioning. Extensive results show that our single model can produce comparable results with specialized methods and better results than prior unified methods. We also demonstrate that multiple models can be effectively combined for multi-signal conditional generation.

3D point cloud object detection plays an important role in autonomous driving. However, labeling 3D object boxes is expensive and time-consuming, limiting the number of annotated point clouds used in fully-supervised training. This has led to a rise in semi-supervised 3D object detection research, which aims to improve model performance by leveraging both labeled and unlabeled point clouds. Existing methods typically rely on the Mean Teacher (MT) paradigm, which uses unlabeled instances discovered by the teacher with confidence scores higher than certain thresholds to train the student. However, this leads to a loss of information as it overlooks ambiguous instances from the teacher that could also contain valuable knowledge. To address this issue, we propose a Bi-Stream Knowledge Transfer (BiKT) framework that fully exploits and transfers knowledge from both confident and ambiguous instances to the student network. Specifically, all pseudo labels are allocated into two knowledge streams, the deterministic stream and the noisy stream, and then subsequently guide the student network through bi-level supervision. We also introduce a Dynamic Stream Switching (DSS) algorithm that sets the stream boundary tailored for the current learning status. To further improve the quality of pseudo labels in the knowledge streams, we propose a Diffusive Label Denoising (DLD) module, which is trained by explicitly generating noised instances and then learning to denoise them, as in diffusion models. Experiments show the state-of-the-art performance of our BiKT on the ONCE validation and testing sets, as well as the robust generalization capability when confronted with diverse base detectors, increased amount of unlabeled data, and distinct datasets (e.g., Waymo), unveiling the power of semi-supervised learning in 3D object detection.

Foundational vision-language models (VLMs) like CLIP have revolutionized image recognition, but adapting them to specialized domains with limited data remains challenging. We propose Domain Prompt Learning with Quaternion Networks (DPLQ), which leverages domain-specific foundation models and quaternion-based prompt tuning to effectively transfer recognition capabilities. Our method achieves state-of-the-art results in remote sensing and medical imaging tasks. This extended abstract highlights the key contributions and performance of DPLQ.

Adversarial patches are widely used to evaluate the robustness of object detection systems in real-world scenarios. These patches were initially designed to deceive single-modal detectors (e.g., visible or infrared) and have recently been extended to target visible-infrared dual-modal detectors. However, existing dual-modal adversarial patch attacks have limited attack effectiveness across diverse physical scenarios. To address this, we propose CDUPatch, a universal cross-modal patch attack against visible-infrared object detectors across scales, views, and scenarios. Specifically, we observe that color variations lead to different levels of thermal absorption, resulting in temperature differences in infrared imaging. Leveraging this property, we propose an RGB-to-infrared adapter that maps RGB patches to infrared patches, enabling unified optimization of cross-modal patches. By learning an optimal color distribution on the adversarial patch, we can manipulate its thermal response and generate an adversarial infrared texture. Additionally, we introduce a multi-scale clipping strategy and construct a new visible-infrared dataset, MSDrone, which contains aerial vehicle images in varying scales and perspectives. These data augmentation strategies enhance the robustness of our patch in real-world conditions. Experiments on four benchmark datasets (e.g., DroneVehicle, LLVIP, VisDrone, MSDrone) show that our method outperforms existing patch attacks in the digital domain. Extensive physical tests further confirm strong transferability across scales, views, and scenarios. Attack demos are provided in the supplementary materials.

Low-Rank Adaptation (LoRA) methods have demonstrated considerable success in achieving parameter-efficient fine-tuning (PEFT) for Transformer-based foundation models. These methods typically fine-tune individual Transformer layers using independent LoRA adaptations. However, directly applying existing LoRA techniques to convolutional networks (ConvNets) yields unsatisfactory results due to the high correlation between the stacked sequential layers of ConvNets. To overcome this challenge, we introduce a novel framework called Correlated Low-Rank Adaptation (CoLoRA), which explicitly utilizes correlated low-rank matrices to model the inter-layer dependencies among convolutional layers. Additionally, to enhance tuning efficiency, we propose a parameter-free filtering method that enlarges the receptive field of LoRA, thus minimizing interference from non-informative local regions. Comprehensive experiments conducted across various mainstream vision tasks, including image classification, semantic segmentation, and object detection, illustrate that CoLoRA significantly advances the state-of-the-art PEFT approaches. Notably, our CoLoRA achieves superior performance with only 5\% of trainable parameters, surpassing full fine-tuning in the image classification task on the VTAB-1k dataset using ConvNeXt-S. Code is available at [https://github.com/VISION-SJTU/CoLoRA](https://github.com/VISION-SJTU/CoLoRA).

The generalization capability of a face anti-spoofing (FAS) model is critical to its practicality in the real world. Recent studies have theoretically and empirically uncovered that neural networks tend to exploit easy-to-learn frequency sets for decisions. These simplicity-biased representations, depending on what best simplifies the training objective, may hamper generalization. This paper thus focuses on mitigating the frequency shortcut learning of prior FAS models for improved generalization. Specifically, we introduce a frequency-aware autoencoder to retain more frequency details in intermediate features via reconstruction, facilitating comprehensive judgment of FAS. Based on the encoder output, we propose a dynamic frequency masking mechanism to select and suppress the probable shortcut bands during training, enabling broader horizons on under-explored frequencies. Moreover, we employ a style inhibited modulation to weaken stylized information in frequency space to reduce the reliance on spurious style features. Experiment results on generalized FAS benchmarks verify the superiority of our framework over existing methods. Our code has been integrated into this project: https://github.com/VISION-SJTU/UniDefense.

Open-vocabulary image semantic segmentation (OVS) seeks to segment images into semantic regions across an open set of categories. Existing OVS methods commonly depend on foundational vision-language models and utilize similarity computation to tackle OVS tasks. However, these approaches are predominantly tailored to natural images and struggle with the unique characteristics of high-resolution remote sensing images, such as rapidly changing orientations and significant scale variations. To tackle this dilemma, we propose the first OVS framework specifically designed for high-resolution remote sensing imagery, introducing a rotation-aggregative similarity computation module to enhance segmentation across varying orientations and a multi-scale feature integration strategy to generate scale-aware semantic masks. Additionally, we establish the first open-sourced OVS benchmark for remote sensing, comprising four public datasets. Experiments demonstrate that our framework effectively addresses orientation and scale challenges, achieving state-of-the-art performance. All codes and datasets are available at https://github.com/caoql98/OVRS.

Pre-trained vision-language models (VLMs), equipped with parameter-efficient tuning (PET) methods like prompting, have shown impressive knowledge transferability on new downstream tasks, but they are still prone to be limited by catastrophic forgetting and overfitting dilemma due to large gaps among tasks. Furthermore, the underlying physical mechanisms of prompt-based tuning methods (especially for visual prompting) remain largely unexplored. It is unclear why these methods work solely based on learnable parameters as prompts for adaptation. To address the above challenges, we present a new prompt-based framework for vision-language models, termed Uni-prompt. Our framework transfers VLMs to downstream tasks by designing visual prompts from an attention perspective that reduces the transfer/solution space, which enables the vision model to focus on task-relevant regions of the input image while also learning task-specific knowledge. Additionally, Uni-prompt further aligns visual-text prompts learning through a pretext task with masked representation modeling interactions, which implicitly learns a global cross-modal matching between visual and language concepts for consistency. We conduct extensive experiments on the few-shot classification task and achieve significant improvement using our Uni-prompt method while requiring minimal extra parameters cost.

Large pre-trained vision-language models, such as CLIP, have shown remarkable generalization capabilities across various tasks when appropriate text prompts are provided. However, adapting these models to specific domains, like remote sensing images (RSIs), medical images, etc, remains unexplored and challenging. Existing prompt learning methods often lack domain-awareness or domain-transfer mechanisms, leading to suboptimal performance due to the misinterpretation of specific images in natural image patterns. To tackle this dilemma, we proposed a Domain-Controlled Prompt Learning for the specific domains. Specifically, the large-scale specific domain foundation model (LSDM) is first introduced to provide essential specific domain knowledge. Using lightweight neural networks, we transfer this knowledge into domain biases, which control both the visual and language branches to obtain domain-adaptive prompts in a directly incorporating manner. Simultaneously, to overcome the existing overfitting challenge, we propose a novel noisy-adding strategy, without extra trainable parameters, to help the model escape the suboptimal solution in a global domain oscillation manner. Experimental results show our method achieves state-of-the-art performance in specific domain image recognition datasets. Our code is available at https://github.com/caoql98/DCPL.

A fundamental task in the realms of computer vision, Low-Rank Matrix Recovery (LRMR) focuses on the inherent low-rank structure precise recovery from incomplete data and/or corrupted measurements given that the rank is a known prior or accurately estimated. However, it remains challenging for existing rank estimation methods to accurately estimate the rank of an ill-conditioned matrix. Also, existing LRMR optimization methods are heavily dependent on the chosen parameters, and are therefore difficult to adapt to different situations. Addressing these issues, A novel LEarning-based low-rank matrix recovery with Rank Estimation (LERE) is proposed. More specifically, considering the characteristics of the Gerschgorin disk's center and radius, a new heuristic decision rule in the Gerschgorin Disk Theorem is significantly enhanced and the low-rank boundary can be exactly located, which leads to a marked improvement in the accuracy of rank estimation. According to the estimated rank, we select row and column sub-matrices from the observation matrix by uniformly random sampling. A 17-iteration feedforward-recurrent-mixed neural network is then adapted to learn the parameters in the sub-matrix recovery processing. Finally, by the correlation of the row sub-matrix and column sub-matrix, LERE successfully recovers the underlying low-rank matrix. Overall, LERE is more efficient and robust than existing LRMR methods. Experimental results demonstrate that LERE surpasses state-of-the-art (SOTA) methods. The code for this work is accessible at https://github.com/zhengqinxu/LERE.

Diffusion models have made significant advances in generating high-quality images, but their application to video generation has remained challenging due to the complexity of temporal motion. Zero-shot video editing offers a solution by utilizing pre-trained image diffusion models to translate source videos into new ones. Nevertheless, existing methods struggle to maintain strict temporal consistency and efficient memory consumption. In this work, we propose a novel approach to enhance temporal consistency in generated videos by merging self-attention tokens across frames. By aligning and compressing temporally redundant tokens across frames, our method improves temporal coherence and reduces memory consumption in self-attention computations. The merging strategy matches and aligns tokens according to the temporal correspondence between frames, facilitating natural temporal consistency in generated video frames. To manage the complexity of video processing, we divide videos into chunks and develop intra-chunk local token merging and inter-chunk global token merging, ensuring both short-term video continuity and long-term content consistency. Our video editing approach seamlessly extends the advancements in image editing to video editing, rendering favorable results in temporal consistency over state-of-the-art methods.

Recent 3D face reconstruction methods have made re-markable advancements, yet there remain huge challenges in monocular high-quality facial reflectance reconstruction. Existing methods rely on a large amount of light-stage captured data to learn facial reflectance models. However, the lack of subject diversity poses challenges in achieving good generalization and widespread applicability. In this paper, we learn the reflectance prior in image space rather than UV space and present a framework named ID2Reflectance. Our framework can directly estimate the reflectance maps of a single image while using limited reflectance data for training. Our key insight is that reflectance data shares facial structures with RGB faces, which enables obtaining expressive facial prior from inexpensive RGB data thus re-ducing the dependency on reflectance data. We first learn a high-quality prior for facial reflectance. Specifically, we pretrain multi-domain facial feature code books and design a codebook fusion method to align the reflectance and RGB domains. Then, we propose an identity-conditioned swapping module that injects facial identity from the target image into the pre-trained autoencoder to modify the identity of the source reflectance image. Finally, we stitch multi-view swapped reflectance images to obtain renderable assets. Extensive experiments demonstrate that our method exhibits excellent generalization capability and achieves state-of-the-art facial reflectance reconstruction results for in-the-wild faces. Our project page is https://xingyuren.github.io/id2reflectance.

Vision-based perception for autonomous driving requires an explicit modeling of a 3D space, where 2D latent representations are mapped and subsequent 3D operators are applied. However, operating on dense latent spaces introduces a cubic time and space complexity, which limits scalability in terms of perception range or spatial resolution. Existing approaches compress the dense representation using projections like Bird's Eye View (BEV) or Tri-Perspective View (TPV). Although efficient, these projections result in information loss, especially for tasks like semantic occupancy prediction. To address this, we propose SparseOcc, an efficient occupancy network inspired by sparse point cloud processing. It utilizes a lossless sparse latent representation with three key innovations. Firstly, a 3D sparse diffuser performs latent completion using spatially decomposed 3D sparse convolutional kernels. Secondly, a feature pyramid and sparse interpolation enhance scales with information from others. Finally, the transformer head is redesigned as a sparse variant. SparseOcc achieves a remarkable 74.9% reduction on FLOPs over the dense baseline. Interestingly, it also improves accuracy, from 12.8% to 14.1% mIOU, which in part can be attributed to the sparse representation's ability to avoid hallucinations on empty voxels.

In the realm of video object tracking, auxiliary modalities such as depth, thermal, or event data have emerged as valuable assets to complement the RGB trackers. In practice, most existing RGB trackers learn a single set of parameters to use them across datasets and applications. However, a similar single-model unification for multi-modality tracking presents several challenges. These challenges stem from the inherent heterogeneity of inputs - each with modality-specific representations, the scarcity of multi-modal datasets, and the absence of all the modalities at all times. In this work, we introduce Un-Track, a Unified Tracker of a single set of parameters for any modality. To handle any modality, our method learns their common latent space through low-rank factorization and reconstruction techniques. More importantly, we use only the RGB-X pairs to learn the common latent space. This unique shared representation seamlessly binds all modalities together, enabling effective unification and accommodating any missing modality, all within a single transformer-based architecture. Our Un-Track achieves +8.1 absolute F-score gain, on the DepthTrack dataset, by introducing only +2.14 (over 21.50) GFLOPs with +6.6M (over 93M) parameters, through a simple yet efficient prompting strategy. Extensive comparisons on five benchmark datasets with different modalities show that Un-Track surpasses both SOTA unified trackers and modality-specific counterparts, validating our effectiveness and practicality. The source code is publicly available at https://thub.com/Zongwei97/UnTrack.

Existing Siamese or transformer trackers commonly pose visual object tracking as a one-shot detection problem, i.e., locating the target object in a single forward evaluation scheme. Despite the demonstrated success, these trackers may easily drift towards distractors with similar appear-ance due to the single forward evaluation scheme lacking self-correction. To address this issue, we cast visual tracking as a point set based denoising diffusion process and propose a novel generative learning based tracker, dubbed Diffusion Track. Our DiffusionTrack possesses two appealing properties: 1) It follows a novel noise-to-target tracking paradigm that leverages multiple denoising diffusion steps to localize the target in a dynamic searching man-ner per frame. 2) It models the diffusion process using a point set representation, which can better handle appear-ance variations for more precise localization. One side benefit is that DiffusionTrack greatly simplifies the post-processing, e.g. removing window penalty scheme. Without bells and whistles, our DiffusionTrack achieves leading per-formance over the state-of-the-art trackers and runs in real-time. The code is in https://github.com/VISION-SJTU/DiffusionTrack.

The proliferation of gigapixel imaging has ushered in unprecedented challenges in object detection and tracking due to the intense computational demands. Previous deep learning approaches, often tailored for megapixel images, fall short in addressing the unique complexities presented by the gigapixel level. To bridge this gap, we introduce SaccadeMOT, a novel architecture designed for efficient gigapixel-level multi-object tracking. Based on our observations of density map regression in crowd counting and small object detection in object detection tasks, we propose a novel gigapixel detection paradigm that combines the strengths of both approaches. Firstly, the “saccade” stage swiftly identifies regions likely containing objects, followed by the “gaze” stage that refines the detection within these areas. This strategic region selection is complemented by a robust tracking mechanism that combines head and body tracking, enhancing accuracy in environments with potential occlusions. Validated on the PANDA dataset, SaccadeMOT not only demonstrates an 13× speed improvement over existing state-of-the-art tracker BotSORT but also exhibits promising applications in gigapixel-level pathology analysis, particularly in Whole Slide Imaging (WSI). This approach sets a new benchmark for handling super high-resolution images, offering significant advancements in both the speed and precision of object tracking technologies.

Multimodal pre-trained models have shown impressive potential in enhancing performance on downstream tasks. However, existing fusion strategies for modalities primarily rely on explicit interaction structures that fail to capture the diverse aspects and patterns inherent in input data. This yields limited performance in zero-shot contexts, especially when fine-grained classifications and abstract interpretations are required. To address this, we propose an effective approach, namely Prompt Learning with Quaternion Networks (QNet), for semantic alignment across diverse modalities. QNet employs a quaternion hidden space where the mutually orthogonal imaginary axes capture rich intermodal semantic spatial correlations from various perspectives. Hierarchical features across multilayers are utilized to encode intricate interdependencies within various modalities with reduced parameters. Our experiments on 11 datasets demonstrate that QNet outperforms state-of-the-art prompt learning techniques in base-to-novel generalization, cross-dataset transfer, and domain transfer scenarios with fewer learnable parameters. The source code is available at https://github.com/VISION-SJTU/QNet.

Recent years have seen an increase in the use of gigapixel-level image and video capture systems and benchmarks with high-resolution wide (HRW) shots. However, unlike close-up shots in the MS COCO dataset, the higher resolution and wider field of view raise unique challenges, such as extreme sparsity and huge scale changes, causing existing close-up detectors inaccuracy and inefficiency. In this paper, we present a novel model-agnostic sparse vision transformer, dubbed SparseFormer, to bridge the gap of object detection between close-up and HRW shots. The proposed SparseFormer selectively uses attentive tokens to scrutinize the sparsely distributed windows that may contain objects. In this way, it can jointly explore global and local attention by fusing coarse- and fine-grained features to handle huge scale changes. SparseFormer also benefits from a novel Cross-slice non-maximum suppression (C-NMS) algorithm to precisely localize objects from noisy windows and a simple yet effective multi-scale strategy to improve accuracy. Extensive experiments on two HRW benchmarks, PANDA and DOTA-v1.0, demonstrate that the proposed SparseFormer significantly improves detection accuracy (up to 5.8%) and speed (up to 3x) over the state-of-the-art approaches.

Knowledge distillation (KD) is an established paradigm for transferring privileged knowledge from a cumbersome model to a more lightweight and efficient one. In recent years, logit-based KD methods are quickly catching up in performance with their feature-based counterparts. However, previous research has pointed out that logit-based methods are still fundamentally limited by two major issues in their training process, namely overconfident teacher and confirmation bias. Inspired by the success of cross-view learning in fields such as semi-supervised learning, in this work we introduce within-view and cross-view regularisations to standard logit-based distillation frameworks to combat the above cruxes. We also perform confidence-based soft label mining to improve the quality of distilling signals from the teacher, which further mitigates the confirmation bias problem. Despite its apparent simplicity, the proposed Consistency-Regularisation-based Logit Distillation (CRLD) significantly boosts student learning, setting new state-of-the-art results on the standard CIFAR-100, Tiny-ImageNet, and ImageNet datasets across a diversity of teacher and student architectures, whilst introducing no extra network parameters. Orthogonal to on-going logit-based distillation research, our method enjoys excellent generalisation properties and, without bells and whistles, boosts the performance of various existing approaches by considerable margins.

While humans effortlessly discern intrinsic dynamics and adapt to new scenarios, modern AI systems often struggle. Current methods for visual grounding of dynamics either use pure neural-network-based simulators (black box), which may violate physical laws, or traditional physical simulators (white box), which rely on expert-defined equations that may not fully capture actual dynamics. We propose the Neural Material Adaptor (NeuMA), which integrates existing physical laws with learned corrections, facilitating accurate learning of actual dynamics while maintaining the generalizability and interpretability of physical priors. Additionally, we propose Particle-GS, a particle-driven 3D Gaussian Splatting variant that bridges simulation and observed images, allowing back-propagate image gradients to optimize the simulator. Comprehensive experiments on various dynamics in terms of grounded particle accuracy, dynamic rendering quality, and generalization ability demonstrate that NeuMA can accurately capture intrinsic dynamics. Project Page: https://xjay18.github.io/projects/neuma.html.

Recent advancements in State Space Models, notably Mamba, have demonstrated superior performance over the dominant Transformer models, particularly in reducing the computational complexity from quadratic to linear. Yet, difficulties in adapting Mamba from language to vision tasks arise due to the distinct characteristics of visual data, such as the spatial locality and adjacency within images and large variations in information granularity across visual tokens. Existing vision Mamba approaches either flatten tokens into sequences in a raster scan fashion, which breaks the local adjacency of images, or manually partition tokens into windows, which limits their long-range modeling and generalization capabilities. To address these limitations, we present a new vision Mamba model, coined QuadMamba, that effectively captures local dependencies of varying granularities via quadtree-based image partition and scan. Concretely, our lightweight quadtree-based scan module learns to preserve the 2D locality of spatial regions within learned window quadrants. The module estimates the locality score of each token from their features, before adaptively partitioning tokens into window quadrants. An omnidirectional window shifting scheme is also introduced to capture more intact and informative features across different local regions. To make the discretized quadtree partition end-to-end trainable, we further devise a sequence masking strategy based on Gumbel-Softmax and its straight-through gradient estimator. Extensive experiments demonstrate that QuadMamba achieves state-of-the-art performance in various vision tasks, including image classification, object detection, instance segmentation, and semantic segmentation. Our code and models will be released.

We present Frankenstein, a diffusion-based framework that can generate semantic-compositional 3D scenes in a single pass. Unlike existing methods that output a single, unified 3D shape, Frankenstein simultaneously generates multiple separated shapes, each corresponding to a semantically meaningful part. The 3D scene information is encoded in one single triplane tensor, from which multiple Signed Distance Function (SDF) fields can be decoded to represent the compositional shapes. During training, an auto-encoder compresses tri-planes into a latent space, and then the denoising diffusion process is employed to approximate the distribution of the compositional scenes. Frankenstein demonstrates promising results in generating room interiors as well as human avatars with automatically separated parts. The generated scenes facilitate many downstream applications, such as part-wise re-texturing, object rearrangement in the room or avatar cloth re-targeting.

Monocular 3D object detection (M3OD) is a significant yet inherently challenging task in autonomous driving due to absence of explicit depth cues in a single RGB image. In this paper, we strive to boost currently underperforming monocular 3D object detectors by leveraging an abundance of unlabelled data via semi-supervised learning. Our proposed ODM3D framework entails cross-modal knowledge distillation at various levels to inject LiDAR-domain knowledge into a monocular detector during training. By identifying foreground sparsity as a main culprit behind existing methods’ suboptimal training, we exploit the precise localisation information embedded in LiDAR points to enable more foreground-attentive and efficient distillation via the proposed BEV occupancy guidance mask, leading to notably improved knowledge transfer and M3OD performance. Besides, motivated by insights into why existing cross-modal GT-sampling techniques fail on our task at hand, we further design a novel cross-modal object-wise data augmentation strategy for effective RGB-LiDAR joint learning. Our method ranks 1stin both KITTI validation and test benchmarks, significantly surpassing all existing monocular methods, supervised or semi-supervised, on both BEV and 3D detection metrics. Code will be released at https://github.com/arcaninez/odm3d.

Fusing geometric cues with visual appearance is an imperative theme for RGB-D indoor semantic segmentation . Existing methods commonly adopt convolutional modules to aggregate multi-modal features, paying little attention to explicitly leveraging the long-range dependencies in feature fusion . Therefore, it is challenging for existing methods to accurately segment objects with large-scale variations. In this paper, we propose a novel transformer-based fusion scheme, named TransD-Fusion, to better model contextualized awareness. Specifically, TransD-Fusion consists of a self-refinement module, a calibration scheme with cross-interaction, and a depth-guided fusion. The objective is to first improve modality-specific features with self- and cross-attention, and then explore the geometric cues to better segment objects sharing a similar visual appearance. Additionally, our transformer fusion benefits from a semantic-aware position encoding which spatially constrains the attention to neighboring pixels . Extensive experiments on RGB-D benchmarks demonstrate that the proposed method performs well over the state-of-the-art methods by large margins.

Developing robust and discriminative appearance models has been a long-standing research challenge in visual object tracking. In the prevalent Siamese-based paradigm, the features extracted by the Siamese-like networks are often insufficient to model the tracked targets and distractor objects, thereby hindering them from being robust and discriminative simultaneously. While most Siamese trackers focus on designing robust correlation operations, we propose a novel single-branch tracking framework inspired by the transformer. Unlike the Siamese-like feature extraction, our tracker deeply embeds cross-image feature correlation in multiple layers of the feature network. By extensively matching the features of the two images through multiple layers, it can suppress non-target features, resulting in target-aware feature extraction. The output features can be directly used to predict target locations without additional correlation steps. Thus, we reformulate the two-branch Siamese tracking as a conceptually simple, fully transformer-based Single-Branch Tracking pipeline, dubbed SBT. After conducting an in-depth analysis of the SBT baseline, we summarize many effective design principles and propose an improved tracker dubbed SuperSBT. SuperSBT adopts a hierarchical architecture with a local modeling layer to enhance shallow-level features. A unified relation modeling is proposed to remove complex handcrafted layer pattern designs. SuperSBT is further improved by masked image modeling pre-training, integrating temporal modeling, and equipping with dedicated prediction heads. Thus, SuperSBT outperforms the SBT baseline by 4.7%,3.0%, and 4.5% AUC scores in LaSOT, TrackingNet, and GOT-10K. Notably, SuperSBT greatly raises the speed of SBT from 37 FPS to 81 FPS. Extensive experiments show that our method achieves superior results on eight VOT benchmarks.

Lane detection is a critical component of autonomous driving systems and has been the subject of extensive research. Unlike object detection, identifying car lanes requires extracting features from multi-scale information since they are slender, sparse, and distributed in the entire image. Unfortunately, many previous works have overlooked this requirement, resulting in less robust lane feature extraction. To address this issue, we propose an attention mechanism that is exceptional at extracting global and long-dependence lane features compared to CNNs. Our attention-based structure, calledattention in attention, explores the relationship between various correlations and mitigates mismatched correlation problems in attention computation. Furthermore, we introduce a novel feature fusion structure in the backbone called the double feature pyramid network, which effectively gathers feature information with various dimensions and enlarges the receptive field. Our network is based on the BEV-LaneDet and achieves impressive performance on the OpenLane dataset. Notably, experiments demonstrate that our method surpasses BEV-LaneDet by 4.4% in terms of F-Score on OpenLane.

Few-shot semantic segmentation (FSS) aims to segment objects of unseen classes in query images with only a few annotated support images. Existing FSS algorithms typically focus on mining category representations from the single-view support to match semantic objects of the single-view query. However, the limited annotated samples render the single-view matching struggle to perceive the varying characteristics of novel objects, which results in a restricted learning space for novel categories and further induces a biased segmentation with demoted parsing performance. To address this challenge, inspired by the semantic transform invariance, this paper proposes a fresh few-shot segmentation framework to break the bias and perform invariant segmentation in a multi-view matching manner. Specifically, original and transform support features from different perspectives with the same semantics are learnable fused to obtain the transform invariance prototype with a stronger category representation ability. Simultaneously, aiming at providing better parsing guidance, the Transform Invariance Guidance Mask Generation (TIGM) module is proposed to integrate prior knowledge from different perspectives. Finally, segmentation predictions from varying views are complementarily merged in the Transform Invariance Semantic Prediction (TISP) module to decide the uncertain area and yield precise segmentation predictions. Extensive experiments on both PASCAL-5i and COCO-20i datasets demonstrate the effectiveness of our approach and show that our method could achieve state-of-the-art performance. Code is available at https://github.com/caoql98/BBD.

Few-shot aerial image semantic segmentation is a challenging task that requires precisely parsing unseen-category objects in query aerial images with limited annotated support aerial images. Formally, category prototypes would be extracted from support samples to segment query images in a pixel-to–pixel matching manner. However, aerial objects in aerial images are often distributed with arbitrary orientations, and varying orientations could cause a dramatic feature change. This unique property of aerial images renders conventional matching manner without consideration of orientations fails to activate same-category objects with different orientations. Furthermore, the oscillation of the confidence scores in existing rotation-insensitive algorithms, engendered by the striking changes of object orientations, often leads to false recognition of lower-scored rotated semantic objects. To tackle these challenges, inspired by the intrinsic rotation invariance in aerial images, we propose a novel few-shot rotation-invariant aerial semantic segmentation network (FRINet) to efficiently segment aerial semantic objects with diverse orientations. Specifically, through extracting orientation-varying yet category-consistent support information, FRINet provides rotation-adaptive matching for each query feature in a feature-aggregation manner. Meanwhile, to encourage consistent predictions for aerial objects with arbitrary orientations, segmentation predictions from different orientations are supervised by the same label and further fused to obtain the final rotation-invariant prediction in a complementary manner. Moreover, aiming at providing a better solution searching space, the backbones are newly pre-trained in the base category to basically boost the segmentation performance. Extensive experiments on the few-shot aerial image semantic segmentation benchmark demonstrate that the proposed FRINet achieves a new state-of-the-art performance. The code is available at https://github.com/caoql98/FRINet.

Visual attention advances object detection by attending neural networks to object representations. While existing methods incorporate empirical modules to empower network attention, we rethink attentive object detection from the network learning perspective in this work. We propose a NEural Attention Learning approach (NEAL) which consists of two parts. During the back-propagation of each training iteration, we first calculate the partial derivatives (a.k.a. the accumulated gradients) of the classification output with respect to the input features. We refine these partial derivatives to obtain attention response maps whose elements reflect the contributions to the final network predictions. Then, we formulate the attention response maps as extra objective functions, which are combined together with the original detection loss to train detectors in an end-to-end manner. In this way, we succeed in learning an attentive CNN model without introducing additional network structures. We apply NEAL to the two-stage object detection frameworks, which are usually composed of a CNN feature backbone, a region proposal network (RPN), and a classifier. We show that the proposed NEAL not only helps the RPN attend to objects but also enables the classifier to pay more attention to the premier positive samples. To this end, the localization (proposal generation) and classification mutually benefit from each other in our proposed method. Extensive experiments on large-scale benchmark datasets, including MS COCO 2017 and Pascal VOC 2012, demonstrate that the proposed NEAL algorithm advances the two-stage object detector over state-of-the-art approaches.

Recent years have witnessed the incredible performance boost of data-driven deep visual object trackers. Despite the success, these trackers require millions of sequential manual labels on videos for supervised training, implying the heavy burden of human annotating. This raises a crucial question: how to train a powerful tracker from abundant videos using limited manual annotations? In this paper, we challenge the conventional belief that frame-by-frame labeling is indispensable, and show that providing a small number of annotated bounding boxes in each video is sufficient for training a strong tracker. To facilitate that, we design a novel SParsely-supervised Object Tracking (SPOT) framework. It regards the sparsely annotated boxes as anchors and progressively explores in the temporal span to discover unlabeled target snapshots. Under the teacher-student paradigm, SPOT leverages the unique transitive consistency inherent in the tracking task as supervision, extracting knowledge from both anchor snapshots and unlabeled target snapshots. We also utilize several effective training strategies, i.e., IoU filtering, asymmetric augmentation, and temporal calibration to further improve the training robustness of SPOT. The experimental results demonstrate that, given less than 5 labels for each video, trackers trained via SPOT perform on par with their fully-supervised counterparts. Moreover, our SPOT exhibits two desirable properties: 1) SPOT enables us to fully exploit large-scale video datasets by efficiently allocating sparse labels to more videos even under a limited labeling budget; 2) when equipped with a target discovery module, SPOT can even learn from purely unlabeled videos for performance gain. We hope this work could inspire the community to rethink the current annotation principles and make a step towards practical label-efficient deep tracking.

Modeling the architecture search process on a supernet and applying a differentiable method to find the importance of architecture are among the leading tools for differentiable neural architectures search (DARTS). One fundamental problem in DARTS is how to discretize or select a single-path architecture from the pretrained one-shot architecture. Previous approaches mainly exploit heuristic or progressive search methods for discretization and selection, which are not efficient and easily trapped by local optimizations. To address these issues, we formulate the task of finding a proper single-path architecture as an architecture game among the edges and operations with the strategies "keep" and "drop" and show that the optimal one-shot architecture is a Nash equilibrium of the architecture game. Then, we propose a novel and effective approach for discretizing and selecting a proper single-path architecture, which is based on extracting the single-path architecture that associates the maximal coefficient of the Nash equilibrium with the strategy "keep" in the architecture game. To further improve the efficiency, we employ a mechanism of entangled Gaussian representation of mini-batches, inspired by the classic Parrondo's paradox. If some mini-batch formed uncompetitive strategies, the entanglement of mini-batches would ensure the games be combined and, thus, turn into strong ones. We conduct extensive experiments on benchmark datasets and demonstrate that our approach is significantly faster than the state-of-the-art progressive discretizing methods while maintaining competitive performance with higher maximum accuracy.

Face swapping is an important research topic in computer vision with wide applications in entertainment and privacy protection. Existing methods directly learn to swap 2D facial images, taking no account of the geometric information of human faces. In the presence of large pose variance between the source and the target faces, there always exist undesirable artifacts on the swapped face. In this paper, we present a novel 3D-aware face swapping method that generates high-fidelity and multi-view-consistent swapped faces from single-view source and target images. To achieve this, we take advantage of the strong geometry and texture prior of 3D human faces, where the 2D faces are projected into the latent space of a 3D generative model. By disentangling the identity and attribute features in the latent space, we succeed in swapping faces in a 3D-aware manner, being robust to pose variations while transferring fine-grained facial details. Extensive experiments demonstrate the superiority of our 3D-aware face swapping framework in terms of visual quality, identity similarity, and multi-view consistency. Code is available at https://lyx0208.github.io/3dSwap.

Recent 3D face reconstruction methods have made significant advances in geometry prediction, yet further cosmetic improvements are limited by lagged albedo because inferring albedo from appearance is an ill-posed problem. Although some existing methods consider prior knowledge from illumination to improve albedo estimation, they still produce a light-skin bias due to racially biased albedo models and limited light constraints. In this paper, we reconsider the relationship between albedo and face attributes and propose a ID2Albedo to directly estimate albedo without constraining illumination. Our key insight is that intrinsic semantic attributes such as race, skin color, and age can be used to constrain the albedo map. We first introduce visual-textual cues and design a semantic loss to supervise facial albedo estimation. Specifically, we pre-define text labels such as race, skin color, age, and wrinkles. Then, we employ the text-image model (CLIP) to compute the similarity between the text and the input image, and assign a pseudo-label to each facial image. We constrain generated albedos in the training phase to have the same attributes as the inputs. In addition, we train a high-quality, unbiased facial albedo generator and utilize the semantic loss to learn the mapping from illumination-robust identity features to the albedo latent codes. Finally, our ID2Albedo is trained in a self-supervised way and outperforms state-of-the-art albedo estimation methods in terms of accuracy and fidelity. It is worth mentioning that our approach has excellent generalizability and fairness, especially on in-the- wild data.

Existing methods mainly handle single weather types. However, the connections of different weather conditions at deep representation level are usually ignored. These connections, if used properly, can generate complementary representations for each other to make up insufficient training data, obtaining positive performance gains and better generalization. In this paper, we focus on the very correlated rain and snow to explore their connections at deep representation level. Because sub-optimal connections may cause negative effect, another issue is that if rain and snow are handled in a multi-task learning way, how to find an optimal connection strategy to simultaneously improve deraining and desnowing performance. To build desired connection, we propose a smart knowledge assignment strategy, called SmartAssign, to optimally assign the knowledge learned from both tasks to a specific one. In order to further enhance the accuracy of knowledge assignment, we propose a novel knowledge contrast mechanism, so that the knowledge assigned to different tasks preserves better uniqueness. The inherited inductive biases usually limit the modelling ability of CNNs, we introduce a novel transformer block to constitute the backbone of our network to effectively combine long-range context dependency and local image details. Extensive experiments on seven benchmark datasets verify that proposed SmartAssign explores effective connection between rain and snow, and improves the performances of both deraining and desnowing apparently. The implementation code will be available at https://gitee.com/mindspore/models/tree/master/research/cv/SmartAssign.

Existing Siamese tracking methods, which are built on pair-wise matching between two single frames, heavily rely on additional sophisticated mechanism to exploit temporal information among successive video frames, hindering them from efficiency and industrial deployments. In this work, we resort to sequence-level target matching that can encode temporal contexts into the spatial features through a neat feedforward video model. Specifically, we adapt the standard video transformer architecture to visual tracking by enabling spatiotemporal feature learning directly from frame-level patch sequences. To better adapt to the tracking task, we carefully blend the spatiotemporal information in the video clips through sequential multi-branch triplet blocks, which formulates a video transformer backbone. Our experimental study compares different model variants, such as tokenization strategies, hierarchical structures, and video attention schemes. Then, we propose a disentangled dual-template mechanism that decouples static and dynamic appearance clues over time, and reduces temporal redundancy in video frames. Extensive experiments show that our method, named as Video Track, achieves state-of-the-art results while running in real-time.

Learning a dense 3D model with fine-scale details from a single facial image is highly challenging and ill-posed. To address this problem, many approaches fit smooth geometries through facial prior while learning details as additional displacement maps or personalized basis. However, these techniques typically require vast datasets of paired multi-view data or 3D scans, whereas such datasets are scarce and expensive. To alleviate heavy data dependency, we present a robust texture-guided geometric detail recovery approach using only a single in-the-wild facial image. Specifically, we inpaint occluded facial parts, generate complete textures, and build an accurate multi-view dataset of the target subject. In order to estimate the detailed geometry, we define an implicit signed distance function and employ a physically-based implicit renderer to reconstruct fine geometric details from the generated multiview images. Our method not only recovers accurate facial details but also decomposes the diffuse and specular albedo, normals and shading components in a self-supervised way. Finally, we register the implicit shape details to a 3D Morphable Model template, which can be used in traditional modeling and rendering pipelines. Extensive experiments demonstrate that the proposed approach can reconstruct impressive facial details from a single image, especially when compared with state-of-the-art methods trained on large datasets.

Knowledge transfer from multi-modal, i.e., LiDAR points and images, to a single LiDAR modal can take advantage of complimentary information from modal-fusion but keep a single modal inference speed, showing a promising direction for point cloud semantic segmentation in autonomous driving. Recent advances in point cloud segmentation distill knowledge from strictly aligned point-pixel fusion features while leaving a large number of unmatched image pixels unexplored and unmatched LiDAR points under-benefited. In this paper, we propose a novel approach, named ProtoTransfer, which not only fully exploits image representations but also transfers the learned multi-modal knowledge to all point cloud features. Specifically, based on the basic multi-modal learning framework, we build up a class-wise prototype bank from the strictly-aligned fusion features and encourage all the point cloud features to learn from the prototypes during model training. Moreover, to exploit the massive unmatched point and pixel features, we use a pseudo-labeling scheme and further accumulate these features into the class-wise prototype bank with a carefully designed fusion strategy. Without bells and whistles, our approach demonstrates superior performance over the published state-of-the-arts on two large-scale benchmarks, i.e., nuScenes and SemanticKITTI, and ranks 2nd on the competitive nuScenes Lidarseg challenge leaderboard.

The explosive growth of image data facilitates the fast development of image processing and computer vision methods for emerging visual applications, meanwhile introducing novel distortions to processed images. This poses a grand challenge to existing blind image quality assessment (BIQA) models, which are weak at adapting to subpopulation shift. Recent work suggests training BIQA methods on the combination of all available human-rated IQA datasets. However, this type of approach is not scalable to a large number of datasets and is cumbersome to incorporate a newly created dataset as well. In this paper, we formulate continual learning for BIQA, where a model learns continually from a stream of IQA datasets, building on what was learned from previously seen data. We first identify five desiderata in the continual setting with three criteria to quantify the prediction accuracy, plasticity, and stability, respectively. We then propose a simple yet effective continual learning method for BIQA. Specifically, based on a shared backbone network, we add a prediction head for a new dataset and enforce a regularizer to allow all prediction heads to evolve with new data while being resistant to catastrophic forgetting of old data. We compute the overall quality score by a weighted summation of predictions from all heads. Extensive experiments demonstrate the promise of the proposed continual learning method in comparison to standard training techniques for BIQA, with and without experience replay. We made the code publicly available at https://github.com/zwx8981/BIQA_CL.

RGB-D saliency detection aims to fuse multi-modal cues to accurately localize salient regions. Existing works often adopt attention modules for feature modeling, with few methods explicitly leveraging fine-grained details to merge with semantic cues. Thus, despite the auxiliary depth information, it is still challenging for existing models to distinguish objects with similar appearances but at distinct camera distances. In this paper, from a new perspective, we propose a novel Hierarchical Depth Awareness network (HiDAnet) for RGB-D saliency detection. Our motivation comes from the observation that the multi-granularity properties of geometric priors correlate well with the neural network hierarchies. To realize multi-modal and multi-level fusion, we first use a granularity-based attention scheme to strengthen the discriminatory power of RGB and depth features separately. Then we introduce a unified cross dual-attention module for multi-modal and multi-level fusion in a coarse-to-fine manner. The encoded multi-modal features are gradually aggregated into a shared decoder. Further, we exploit a multi-scale loss to take full advantage of the hierarchical information. Extensive experiments on challenging benchmark datasets demonstrate that our HiDAnet performs favorably over the state-of-the-art methods by large margins. The source code can be found in https://github.com/Zongwei97/HIDANet/.

Recent advances in 3D object detection typically learn voxel-based or point-based representations on point clouds. Point-based methods preserve precise point positions but incur high computational load, whereas voxel-based methods rasterize unordered points into voxel grids efficiently but give rise to an accuracy bottleneck. To take advantage of voxel- and point-based representations, we develop an effective and efficient 3D object detector via a novel voxel-point geometry abstraction scheme. Our motivation is to use coarse voxel representation to accelerate proposal generation while using precise point representation to facilitate proposal refinement. For voxel representation learning, we propose a context enrichment module with a novel 3D sparse interpolation layer to augment raw points with multi-scale context. We further develop a point-based RoI pooling module with explicit position augmentation for proposal refinement. Extensive experiments on the widely used KITTI Dataset and the latest Waymo Open Dataset show that the proposed algorithm outperforms state-of-the-art point-voxel-based methods while running at 24 FPS on the TITAN XP GPU.

Existing face forgery detectors mainly focus on specific forgery patterns like noise characteristics, local textures, or frequency statistics for forgery detection. This causes specialization of learned representations to known forgery patterns presented in the training set, and makes it difficult to detect forgeries with unknown patterns. In this paper, from a new perspective, we propose a forgery detection frame-work emphasizing the common compact representations of genuine faces based on reconstruction-classification learning. Reconstruction learning over real images enhances the learned representations to be aware of forgery patterns that are even unknown, while classification learning takes the charge of mining the essential discrepancy between real and fake images, facilitating the understanding of forgeries. To achieve better representations, instead of only using the encoder in reconstruction learning, we build bipartite graphs over the encoder and decoder features in a multi-scale fashion. We further exploit the reconstruction difference as guidance of forgery traces on the graph output as the final representation, which is fed into the classifier for forgery detection. The reconstruction and classification learning is optimized end-to-end. Extensive experiments on large-scale benchmark datasets demonstrate the superiority of the proposed method over state of the arts.

Various facial manipulation techniques have drawn seri-ous public concerns in morality, security, and privacy. Al- though existing face forgery classifiers achieve promising performance on detecting fake images, these methods are vulnerable to adversarial examples with injected impercep- tible perturbations on the pixels. Meanwhile, many face forgery detectors always utilize the frequency diversity be-tween real and fake faces as a crucial clue. In this paper, in- stead of injecting adversarial perturbations into the spatial domain, we propose a frequency adversarial attack method against face forgery detectors. Concretely, we apply dis-crete cosine transform (DCT) on the input images and in-troduce a fusion module to capture the salient region of ad-versary in the frequency domain. Compared with existing adversarial attacks (e.g. FGSM, PGD) in the spatial do-main, our method is more imperceptible to human observers and does not degrade the visual quality of the original images. Moreover, inspired by the idea of meta-learning, we also propose a hybrid adversarial attack that performs at-tacks in both the spatial and frequency domains. Exten-sive experiments indicate that the proposed method fools not only the spatial-based detectors but also the state-of- the-art frequency-based detectors effectively. In addition, the proposed frequency attack enhances the transferability across face forgery detectors as black-box attacks.

LiDAR and camera are two common sensors to collect data in time for 3D object detection under the autonomous driving context. Though the complementary information across sensors and time has great potential of benefiting 3D perception, taking full advantage of sequential cross-sensor data still remains challenging. In this paper, we propose a novel LiDAR Image Fusion Transformer (LIFT) to model the mutual interaction relationship of cross-sensor data over time. LIFT learns to align the input 4D sequential cross-sensor data to achieve multi-frame multi-modal information aggregation. To alleviate computational load, we project both point clouds and images into the bird-eye-view maps to compute sparse grid-wise self-attention. LIFT also benefits from a cross-sensor and cross-time data augmentation scheme. We evaluate the proposed approach on the challenging nuScenes and Waymo datasets, where our LIFT performs well over the state-of-the-art and strong baselines.

Deep learning models have shown their vulnerability when dealing with adversarial attacks. Existing attacks almost perform on low-level instances, such as pixels and super-pixels, and rarely exploit semantic clues. For face recognition attacks, existing methods typically generate the l_p-norm perturbations on pixels, however, resulting in low attack transferability and high vulnerability to denoising defense models. In this work, instead of performing perturbations on the low-level pixels, we propose to generate attacks through perturbing on the high-level semantics to improve attack transferability. Specifically, a unified flexible framework, Adversarial Attributes (Adv-Attribute), is designed to generate inconspicuous and transferable attacks on face recognition, which crafts the adversarial noise and adds it into different attributes based on the guidance of the difference in face recognition features from the target. Moreover, the importance-aware attribute selection and the multi-objective optimization strategy are introduced to further ensure the balance of stealthiness and attacking strength. Extensive experiments on the FFHQ and CelebA-HQ datasets show that the proposed Adv-Attribute method achieves the state-of-the-art attacking success rates while maintaining better visual effects against recent attack methods.

Learning to localize sounding objects in visual scenes without manual annotations has drawn increasing attention recently. In this paper, we propose an unsupervised sounding object localization algorithm by using bottom-up and top-down attention in visual scenes. The bottom-up attention module generates an objectness confidence map, while the top-down attention draws the similarity between sound and visual regions. Moreover, we propose a bottom-up attention loss function, which models the correlation relationship between bottom-up and top-down attention. Extensive experimental results demonstrate that our proposed unsupervised method significantly advances the state-of-the-art unsupervised methods. The source code is available at https://github.com/VISION-SJTU/USOL.

In this paper, we address the issue of data imbalance in learning deep models for visual object tracking. Although it is well known that data distribution plays a crucial role in learning and inference models, considerably less attention has been paid to data imbalance in visual tracking. For the deep regression trackers that directly learn a dense mapping from input images of target objects to soft response maps, we identify their performance is limited by the extremely imbalanced pixel-to-pixel differences when computing regression loss. This prevents existing end-to-end learnable deep regression trackers from performing as well as discriminative correlation filters (DCFs) trackers. For the deep classification trackers that draw positive and negative samples to learn discriminative classifiers, there exists heavy class imbalance due to a limited number of positive samples when compared to the number of negative samples. To balance training data, we propose a novel shrinkage loss to penalize the importance of easy training data mostly coming from the background, which facilitates both deep regression and classification trackers to better distinguish target objects from the background. We extensively validate the proposed shrinkage loss function on six benchmark datasets, including the OTB-2013, OTB-2015, UAV-123, VOT-2016, VOT-2018 and LaSOT. Equipped with our shrinkage loss, the proposed one-stage deep regression tracker achieves favorable results against state-of-the-art methods, especially in comparison with DCFs trackers. Meanwhile, our shrinkage loss generalizes well to deep classification trackers. When replacing the original binary cross entropy loss with our shrinkage loss, three representative baseline trackers achieve large performance gains, even setting new state-of-the-art results.

Recent RGBD-based models for saliency detection have attracted research attention. The depth clues such as boundary clues, surface normal, shape attribute, etc., contribute to the identification of salient objects with complicated scenarios. However, most RGBD networks require multi-modalities from the input side and feed them separately through a two-stream design, which inevitably results in extra costs on depth sensors and computation. To tackle these inconveniences, we present in this paper a novel fusion design named modality-guided subnetwork (MGSnet). It has the following superior designs: 1) Our model works for both RGB and RGBD data, and dynamically estimates depth if not available. Taking the inner workings of depth-prediction networks into account, we propose to estimate the pseudo-geometry maps from RGB input — essentially mimicking the multi-modality input. 2) Our MGSnet for RGB SOD results in real-time inference but achieves state-of-the-art performance compared to other RGB models. 3) The flexible and lightweight design of MGS facilitates the integration into RGBD two-streaming models. The introduced fusion design enables a cross-modality interaction to enable further progress but with a minimal cost.

Existing deep deraining models are mainly learned via directly minimizing the statistical differences between rainy images and rain-free ground truths. They emphasize learning a mapping from rainy images to rain-free images with supervision. Despite the demonstrated success, these methods do not perform well on restoring the fine-grained local details or removing blurry rainy traces. In this work, we aim to exploit the intrinsic priors of rainy images and develop intrinsic loss functions to facilitate training deraining networks, which decompose a rainy image into a rain-free background layer and a rainy layer containing intact rain streaks. To this end, we introduce the quasi-sparsity prior to train network so as to generate two sparse layers with intact textures of different objects. Then we explore the low-value prior to compensate sparsity, forcing all rain streaks to enter into one layer while non-rain con-tents into another layer to restore image details. We introduce a multi-decoding structure to specially supervise the generation of multi-type deraining features. This helps to learn the most contributory features to deraining in respective spaces. Moreover, our model stabilizes the feature values from multi-spaces via information sharing to alleviate potential artifacts, which also accelerates the running speed. Extensive experiments show that the proposed de-raining method outperforms the state-of-the-art approaches in terms of effectiveness and efficiency.

Adversarial attack arises due to the vulnerability of deep neural networks to perceive input samples injected with imperceptible perturbations. Recently, adversarial attack has been applied to visual object tracking to evaluate the robustness of deep trackers. Assuming that the model structures of deep trackers are known, a variety of white-box attack approaches to visual tracking have demonstrated promising results. However, the model knowledge about deep trackers is usually unavailable in real applications. In this paper, we propose a decision-based black-box attack method for visual object tracking. In contrast to existing black-box adversarial attack methods that deal with static images for image classification, we propose IoU attack that sequentially generates perturbations based on the predicted IoU scores from both current and historical frames. By decreasing the IoU scores, the proposed attack method degrades the accuracy of temporal coherent bounding boxes (i.e., object motions) accordingly. In addition, we transfer the learned perturbations to the next few frames to initialize temporal motion attack. We validate the proposed IoU attack on state-of-the-art deep trackers (i.e., detection based, correlation filter based, and long-term trackers). Extensive experiments on the benchmark datasets indicate the effectiveness of the proposed IoU attack method. The source code is available at https://github.com/VISION-SJTU/IoUattack.

Camera and LiDAR are two complementary sensors for 3D object detection in the autonomous driving context. Camera provides rich texture and color cues while LiDAR specializes in relative distance sensing. The challenge of 3D object detection lies in effectively fusing 2D camera images with 3D LiDAR points. In this paper, we present a novel cross-modal 3D object detection algorithm, named PointAugmenting. On one hand, PointAugmenting decorates point clouds with corresponding point-wise CNN features extracted by pretrained 2D detection models, and then performs 3D object detection over the decorated point clouds. In comparison with highly abstract semantic segmentation scores to decorate point clouds, CNN features from detection networks adapt to object appearance variations, achieving significant improvement. On the other hand, PointAugmenting benefits from a novel cross-modal data augmentation algorithm, which consistently pastes virtual objects into images and point clouds during network training. Extensive experiments on the large-scale nuScenes and Waymo datasets demonstrate the effectiveness and efficiency of our PointAugmenting. Notably, PointAugmenting outperforms the LiDAR-only baseline detector by +6.5% mAP and achieves the new state-of-the-art results on the nuScenes leaderboard to date.

In this paper, we propose to learn an Unsupervised Single Object Tracker (USOT) from scratch. We identify that three major challenges, i.e., moving object discovery, rich temporal variation exploitation, and online update, are the central causes of the performance bottleneck of existing unsupervised trackers. To narrow the gap between unsupervised trackers and supervised counterparts, we propose an effective unsupervised learning approach composed of three stages. First, we sample sequentially moving objects with unsupervised optical flow and dynamic programming, instead of random cropping. Second, we train a naive Siamese tracker from scratch using single-frame pairs. Third, we continue training the tracker with a novel cycle memory learning scheme, which is conducted in longer temporal spans and also enables our tracker to update online. Extensive experiments show that the proposed USOT learned from unlabeled videos performs well over the state-of-the-art unsupervised trackers by large margins, and on par with recent supervised deep trackers. Code is available at https://github.com/VISION-SJTU/USOT.

Detecting and tracking objects in 3D scenes play crucial roles in autonomous driving. Successfully recognizing objects through space and time hinges on a strong detector and a reliable association scheme. Recent 3D detection and tracking approaches widely represent objects as points when associating detection results with trajectories. Despite the demonstrated success, these approaches do not fully exploit the rich appearance information of objects. In this paper, we present a conceptually simple yet effective algorithm, named AlphaTrack, which considers both the location and appearance changes to perform joint 3D object detection and tracking. To achieve this, we propose a cross-modal fusion scheme that fuses camera appearance feature with LiDAR feature to facilitate 3D detection and tracking. We further attach an additional branch to the 3D detector to output instance-aware appearance embedding, which significantly improves tracking performance with our designed association mechanisms. Extensive validations on large-scale autonomous driving dataset demonstrate the effectiveness of the proposed algorithm in comparison with state-of-the-art approaches. Notably, the proposed algorithm ranks first on the nuScenes tracking leaderboard to date.

Most existing point cloud detection models require large-scale, densely annotated datasets. They typically underperform in domain adaptation settings, due to geometry shifts caused by different physical environments or LiDAR sensor configurations. Therefore, it is challenging but valuable to learn transferable features between a labeled source domain and a novel target domain, without any access to target labels. To tackle this problem, we introduce the framework of 3D Contrastive Co-training (3D-CoCo) with two technical contributions. First, 3D-CoCo is inspired by our observation that the bird-eye-view (BEV) features are more transferable than low-level geometry features. We thus propose a new co-training architecture that includes separate 3D encoders with domain-specific parameters, as well as a BEV transformation module for learning domain-invariant features. Second, 3D-CoCo extends the approach of contrastive instance alignment to point cloud detection, whose performance was largely hindered by the mismatch between the fictitious distribution of BEV features, induced by pseudo-labels, and the true distribution. The mismatch is greatly reduced by 3D-CoCo with transformed point clouds, which are carefully designed by considering specific geometry priors. We construct new domain adaptation benchmarks using three large-scale 3D datasets. Experimental results show that our proposed 3D-CoCo effectively closes the domain gap and outperforms the state-of-the-art methods by large margins.

In this paper, we focus on exploring the fusion of images and point clouds for 3D object detection in view of the complementary nature of the two modalities, i.e., images possess more semantic information while point clouds specialize in distance sensing. To this end, we present a novel two-stage multi-modal fusion network for 3D object detection, taking both binocular images and raw point clouds as input. The whole architecture facilitates two-stage fusion. The first stage aims at producing 3D proposals through sparse pointwise feature fusion. Within the first stage, we further exploit a joint anchor mechanism that enables the network to utilize 2D-3D classification and regression simultaneously for better proposal generation. The second stage works on the 2D and 3D proposal regions and fuses their dense features. In addition, we propose to use pseudo LiDAR points from stereo matching as a data augmentation method to densify the LiDAR points, as we observe that objects missed by the detection network mostly have too few points especially for far-away objects. Our experiments on the KITTI dataset show that the proposed multi-stage fusion helps the network to learn better representations.

In this paper, we propose an adaptive region proposal scheme with feature channel regularization to facilitate robust object tracking. We consider tracking as a linear regression problem and an ensemble of correlation filters is trained on-line to distinguish the foreground target from the background. Further, we integrate adaptively learned region proposals into an enhanced two-stream tracking framework based on correlation filters. For the tracking stream, we learn two-stage cascade correlation filters on deep convolutional features to ensure competitive tracking performance. For the detection stream, we employ adaptive region proposals, which are effective in recovering target objects from tracking failures caused by heavy occlusion or out-of-view movement. In contrast to traditional tracking-by-detection methods using random samples or sliding windows, we perform target re-detection over adaptively learned region proposals. Since region proposals naturally take the objectness information into account, we show that the proposed adaptive region proposals can handle the challenging scale estimation problem as well. In addition, we observe the channel redundancy and noisy of feature representation, especially for the convolutional features. Thus, we apply a channel regularization to the correlation filter learning. Extensive experimental validations on OTB, VOT and UAV-123 datasets demonstrate that the proposed method performs favorably against state-of-the-art tracking algorithms.

By bundling multiple complex sub-problems into a unified framework, end-to-end deep learning frameworks reduce the need for hand engineering or tuning of parameters for each component, and optimize different modules jointly to ensure the generalization of the whole deep architecture. Despite tremendous success in numerous computer vision tasks, end-to-end learnings for multi-object tracking (MOT), especially for the assignment problem in data association, have been surprisingly less investigated mainly due to the lack of available training data. Furthermore, it is challenging to discriminate target objects under mutual occlusions or to reduce identity switches in crowded scenes. To tackle these challenges, this paper proposes learning deep conditional random field (CRF) networks, aiming to model the assignment costs as unary potentials and the long-term dependencies among detection results as pairwise potentials. Specifically, we use a bidirectional long short-term memory (LSTM) network to encode the long-term dependencies. We pose the CRF inference as a recurrent neural network learning process using the standard gradient descent algorithm, where unary and pairwise potentials are jointly optimized in an end-to-end manner. Extensive experiments are conducted on the challenging MOT datasets including MOT15, MOT16 and MOT17, and the results show that the proposed algorithm performs favorably against the state-of-the-art methods.

In the above article[1], the title is incorrect. The correct title is “End-to-End Learning Deep CRF Models for Multi-Object Tracking.”

The emerging vision-and-language navigation (VLN) problem aims at learning to navigate an agent to the target location in unseen photo-realistic environments according to the given language instruction. The main challenges of VLN arise mainly from two aspects: first, the agent needs to attend to the meaningful paragraphs of the language instruction corresponding to the dynamically-varying visual environments; second, during the training process, the agent usually imitate the expert demonstrations, i.e., the shortest-path to the target location specified by associated language instructions. Due to the discrepancy of action selection between training and inference, the agent solely on the basis of imitation learning does not perform well. Existing VLN approaches address this issue by sampling the next action from its predicted probability distribution during the training process. This allows the agent to explore diverse routes from the environments, yielding higher success rates. Nevertheless, without being presented with the golden shortest navigation paths during the training process, the agent may arrive at the target location through an unexpected longer route. To overcome these challenges, we design a cross-modal grounding module, which is composed of two complementary attention mechanisms, to equip the agent with a better ability to track the correspondence between the textual and visual modalities. We then propose to recursively alternate the learning schemes of imitation and exploration to narrow the discrepancy between training and inference. We further exploit the advantages of both these two learning schemes via adversarial learning. Extensive experimental results on the Room-to-Room (R2R) benchmark dataset demonstrate that the proposed learning scheme is generalized and complementary to prior arts. Our method performs well against state-of-the-art approaches in terms of effectiveness and efficiency.

Facilitated by deep neural networks, numerous tracking methods have made significant advances. Existing deep trackers mainly utilize independent frames to model the target appearance, while paying less attention to its temporal coherence. In this paper, we propose a recurrent memory activation network (RMAN) to exploit the untapped temporal coherence of the target appearance for visual tracking. We build the RMAN on top of the long short-term memory network (LSTM) with an additional memory activation layer. Specifically, we first use the LSTM to model the temporal changes of the target appearance. Then we selectively activate the memory blocks via the activation layer to produce a temporally coherent representation. The recurrent memory activation layer enriches the target representations from independent frames and reduces the background interference through temporal consistency. The proposed RMAN is fully differentiable and can be optimized end-to-end. To facilitate network training, we propose a temporal coherence loss together with the original binary classification loss. Extensive experimental results on standard benchmarks demonstrate that our method performs favorably against the state-of-the-art approaches.

Existing tracking-by-detection approaches using deep features have achieved promising results in recent years. However, these methods mainly exploit feature representations learned from individual static frames, thus paying little attention to the temporal smoothness between frames. This easily leads trackers to drift in the presence of large appearance variations and occlusions. To address this issue, we propose a two-stream network to learn discriminative spatio-temporal feature representations to represent the target objects. The proposed network consists of a Spatial ConvNet module and a Temporal ConvNet module. Specifically, the Spatial ConvNet adopts 2D convolutions to encode the target-specific appearance in static frames, while the Temporal ConvNet models the temporal appearance variations using 3D convolutions and learns consistent temporal patterns in a short video clip. Then we propose a proposal refinement module to adjust the predicted bounding box, which can make the target localizing outputs to be more consistent in video sequences. In addition, to improve the model adaptation during online update, we propose a contrastive online hard example mining (OHEM) strategy, which selects hard negative samples and enforces them to be embedded in a more discriminative feature space. Extensive experiments conducted on the OTB, Temple Color and VOT benchmarks demonstrate that the proposed algorithm performs favorably against the state-of-the-art methods.

Recently, deep Siamese matching networks have attracted increasing attention for visual tracking. Despite the demonstrated successes, Siamese trackers do not take full advantage of the structural information of target objects. They tend to drift in the presence of non-rigid deformation or partly occlusion. In this paper, we propose to advance Siamese trackers with graph convolutional networks, which pay more attention to the structural layout of target objects, to learn features robust to large appearance changes over time. Specifically, we divide the target object into several sub-parts and design an attentive graph convolutional network to model the relationship between parts. We incrementally update the attention coefficients of the graph with the attention scheme at each frame in an end-to-end manner. To further improve localization accuracy, we propose a learnable cascade regression algorithm based on deep reinforcement learning to refine the predicted bounding boxes. Extensive experiments on seven challenging benchmark datasets, i.e., OTB-100, TC-128, VOT2018, VOT2019, TrackingNet, GOT-10k and LaSOT, demonstrate that the proposed tracking method performs favorably against state-of-the-art approaches.

Correlation filters (CF) have received considerable attention in visual tracking because of their computational efficiency. Leveraging deep features via off-the-shelf CNN models (e.g., VGG), CF trackers achieve state-of-the-art performance while consuming a large number of computing resources. This limits deep CF trackers to be deployed to many mobile platforms on which only a single-core CPU is available. In this paper, we propose to jointly compress and transfer off-the-shelf CNN models within a knowledge distillation framework. We formulate a CNN model pretrained from the image classification task as a teacher network, and distill this teacher network into a lightweight student network as the feature extractor to speed up CF trackers. In the distillation process, we propose a fidelity loss to enable the student network to maintain the representation capability of the teacher network. Meanwhile, we design a tracking loss to adapt the objective of the student network from object recognition to visual tracking. The distillation process is performed offline on multiple layers and adaptively updates the student network using a background-aware online learning scheme. The online adaptation stage exploits the background contents to improve the feature discrimination of the student network. Extensive experiments on six standard datasets demonstrate that the lightweight student network accelerates the speed of state-of-the-art deep CF trackers to real-time on a single-core CPU while maintaining almost the same tracking accuracy.

Video frame interpolation aims to synthesize nonexistent frames in-between the original frames. While significant advances have been made from the recent deep convolutional neural networks, the quality of interpolation is often reduced due to large object motion or occlusion. In this work, we propose a video frame interpolation method which explicitly detects the occlusion by exploring the depth information. Specifically, we develop a depth-aware flow projection layer to synthesize intermediate flows that preferably sample closer objects than farther ones. In addition, we learn hierarchical features to gather contextual information from neighboring pixels. The proposed model then warps the input frames, depth maps, and contextual features based on the optical flow and local interpolation kernels for synthesizing the output frame. Our model is compact, efficient, and fully differentiable. Quantitative and qualitative results demonstrate that the proposed model performs favorably against state-of-the-art frame interpolation methods on a wide variety of datasets. The source code and pre-trained model are available at https://github.com/baowenbo/DAIN.

Existing deep trackers mainly use convolutional neural networks pre-trained for the generic object recognition task for representations. Despite demonstrated successes for numerous vision tasks, the contributions of using pre-trained deep features for visual tracking are not as significant as that for object recognition. The key issue is that in visual tracking the targets of interest can be arbitrary object class with arbitrary forms. As such, pre-trained deep features are less effective in modeling these targets of arbitrary forms for distinguishing them from the background. In this paper, we propose a novel scheme to learn target-aware features, which can better recognize the targets undergoing significant appearance variations than pre-trained deep features. To this end, we develop a regression loss and a ranking loss to guide the generation of target-active and scale-sensitive features. We identify the importance of each convolutional filter according to the back-propagated gradients and select the target-aware features based on activations for representing the targets. The target-aware features are integrated with a Siamese matching network for visual tracking. Extensive experimental results show that the proposed algorithm performs favorably against the state-of-the-art methods in terms of accuracy and speed.

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.

We propose an unsupervised visual tracking method in this paper. Different from existing approaches using extensive annotated data for supervised learning, our CNN model is trained on large-scale unlabeled videos in an unsupervised manner. Our motivation is that a robust tracker should be effective in both the forward and backward predictions (i.e., the tracker can forward localize the target object in successive frames and backtrace to its initial position in the first frame). We build our framework on a Siamese correlation filter network, which is trained using unlabeled raw videos. Meanwhile, we propose a multiple-frame validation method and a cost-sensitive loss to facilitate unsupervised learning. Without bells and whistles, the proposed unsupervised tracker achieves the baseline accuracy of fully supervised trackers, which require complete and accurate labels during training. Furthermore, unsupervised framework exhibits a potential in leveraging unlabeled or weakly labeled data to further improve the tracking accuracy.

Visual tracking is challenging as target objects often undergo significant appearance changes caused by deformation, abrupt motion, background clutter and occlusion. In this paper, we propose to exploit the rich hierarchical features of deep convolutional neural networks to improve the accuracy and robustness of visual tracking. Deep neural networks trained on object recognition datasets consist of multiple convolutional layers. These layers encode target appearance with different levels of abstraction. For example, the outputs of the last convolutional layers encode the semantic information of targets and such representations are invariant to significant appearance variations. However, their spatial resolutions are too coarse to precisely localize the target. In contrast, features from earlier convolutional layers provide more precise localization but are less invariant to appearance changes. We interpret the hierarchical features of convolutional layers as a nonlinear counterpart of an image pyramid representation and explicitly exploit these multiple levels of abstraction to represent target objects. Specifically, we learn adaptive correlation filters on the outputs from each convolutional layer to encode the target appearance. We infer the maximum response of each layer to locate targets in a coarse-to-fine manner. To further handle the issues with scale estimation and re-detecting target objects from tracking failures caused by heavy occlusion or out-of-the-view movement, we conservatively learn another correlation filter, that maintains a long-term memory of target appearance, as a discriminative classifier. We apply the classifier to two types of object proposals: (1) proposals with a small step size and tightly around the estimated location for scale estimation; and (2) proposals with large step size and across the whole image for target re-detection. Extensive experimental results on large-scale benchmark datasets show that the proposed algorithm performs favorably against the state-of-the-art tracking methods.

In this paper, we exploit memory-augmented neural networks to predict accurate answers to visual questions, even when those answers rarely occur in the training set. The memory network incorporates both internal and external memory blocks and selectively pays attention to each training exemplar. We show that memory-augmented neural networks are able to maintain a relatively long-term memory of scarce training exemplars, which is important for visual question answering due to the heavy-tailed distribution of answers in a general VQA setting. Experimental results in two large-scale benchmark datasets show the favorable performance of the proposed algorithm with the comparison to state of the art.

The tracking-by-detection framework consists of two stages, i.e., drawing samples around the target object in the first stage and classifying each sample as the target object or as background in the second stage. The performance of existing trackers using deep classification networks is limited by two aspects. First, the positive samples in each frame are highly spatially overlapped, and they fail to capture rich appearance variations. Second, there exists extreme class imbalance between positive and negative samples. This paper presents the VITAL algorithm to address these two problems via adversarial learning. To augment positive samples, we use a generative network to randomly generate masks, which are applied to adaptively dropout input features to capture a variety of appearance changes. With the use of adversarial learning, our network identifies the mask that maintains the most robust features of the target objects over a long temporal span. In addition, to handle the issue of class imbalance, we propose a high-order cost sensitive loss to decrease the effect of easy negative samples to facilitate training the classification network. Extensive experiments on benchmark datasets demonstrate that the proposed tracker performs favorably against state-of-the-art approaches.

Visual attention, derived from cognitive neuroscience, facilitates human perception on the most pertinent subset of the sensory data. Recently, significant efforts have been made to exploit attention schemes to advance computer vision systems. For visual tracking, it is often challenging to track target objects undergoing large appearance changes. Attention maps facilitate visual tracking by selectively paying attention to temporal robust features. Existing tracking-by-detection approaches mainly use additional attention modules to generate feature weights as the classifiers are not equipped with such mechanisms. In this paper, we propose a reciprocative learning algorithm to exploit visual attention for training deep classifiers. The proposed algorithm consists of feed-forward and backward operations to generate attention maps, which serve as regularization terms coupled with the original classification loss function for training. The deep classifier learns to attend to the regions of target objects robust to appearance changes. Extensive experiments on large-scale benchmark datasets show that the proposed attentive tracking method performs favorably against the state-of-the-art approaches.

We introduce a Multiple Granularity Analysis framework for video segmentation in a coarse-to-fine manner. We cast video segmentation as a spatio-temporal superpixel labeling problem. Benefited from the bounding volume provided by off-the-shelf object trackers, we estimate the foreground/ background super-pixel labeling using the spatiotemporal multiple instance learning algorithm to obtain coarse foreground/background separation within the volume. We further refine the segmentation mask in the pixel level using the graph-cut model. Extensive experiments on benchmark video datasets demonstrate the superior performance of the proposed video segmentation algorithm.

Discriminative correlation filters (DCFs) have been shown to perform superiorly in visual tracking. They only need a small set of training samples from the initial frame to generate an appearance model. However, existing DCFs learn the filters separately from feature extraction, and update these filters using a moving average operation with an empirical weight. These DCF trackers hardly benefit from the end-to-end training. In this paper, we propose the CREST algorithm to reformulate DCFs as a one-layer convolutional neural network. Our method integrates feature extraction, response map generation as well as model update into the neural networks for an end-to-end training. To reduce model degradation during online update, we apply residual learning to take appearance changes into account. Extensive experiments on the benchmark datasets demonstrate that our CREST tracker performs favorably against state-of-the-art trackers.

Correlation filters have been widely applied to visual tracking in recent years as adaptive correlation filters with short-term memory are robust to large appearance changes. However, tracking methods relying on correlation filters are prone to drifting due to noisy updates. Moreover, these methods are unable to recover from tracking failures caused by temporary or persistent heavy occlusions. In this paper, we interpret correlation filters as the counterparts of convolution filters in deep neural networks. Correlation filters encode the holistic template of target appearance, while convolution filters with smaller size encode the part-based template. In the light of this idea, we propose to exploit deep convolutional networks that directly learn mapping as a spatial correlation between two consecutive frames for visual tracking. We show that these deeply learned networks are effective in maintaining the long-term memory of target appearance for handling heavy occlusion or out-of-view. We further take the response maps both from the deep networks and conventional correlation filters into account for precisely locating the target. Experimental results on large-scale benchmark sequences show that the proposed algorithm performs favorably against the state-of-the-art methods.

In this paper, we address the problem of long-term visual tracking where the target objects undergo significant appearance variation due to deformation, abrupt motion, heavy occlusion and out-of-view. In this setting, we decompose the task of tracking into translation and scale estimation of objects. We show that the correlation between temporal context considerably improves the accuracy and reliability for translation estimation, and it is effective to learn discriminative correlation filters from the most confident frames to estimate the scale change. In addition, we train an online random fern classifier to re-detect objects in case of tracking failure. Extensive experimental results on large-scale benchmark datasets show that the proposed algorithm performs favorably against state-of-the-art methods in terms of efficiency, accuracy, and robustness.

Visual object tracking is challenging as target objects often undergo significant appearance changes caused by deformation, abrupt motion, background clutter and occlusion. In this paper, we exploit features extracted from deep convolutional neural networks trained on object recognition datasets to improve tracking accuracy and robustness. The outputs of the last convolutional layers encode the semantic information of targets and such representations are robust to significant appearance variations. However, their spatial resolution is too coarse to precisely localize targets. In contrast, earlier convolutional layers provide more precise localization but are less invariant to appearance changes. We interpret the hierarchies of convolutional layers as a nonlinear counterpart of an image pyramid representation and exploit these multiple levels of abstraction for visual tracking. Specifically, we adaptively learn correlation filters on each convolutional layer to encode the target appearance. We hierarchically infer the maximum response of each layer to locate targets. Extensive experimental results on a largescale benchmark dataset show that the proposed algorithm performs favorably against state-of-the-art methods.

In this paper, we propose to learn temporally invariant features from a large number of image sequences to represent objects for visual tracking. These features are trained on a convolutional neural network with temporal invariance constraints and robust to diverse motion transformations. We employ linear correlation filters to encode the appearance templates of targets and perform the tracking task by searching for the maximum responses at each frame. The learned filters are updated online and adapt to significant appearance changes during tracking. Extensive experimental results on challenging sequences show that the proposed algorithm performs favorably against state-of-the-art methods in terms of efficiency, accuracy, and robustness.

Sketch retrieval aims at retrieving most similar sketches from a large database based on one hand-drawn query. Successful retrieval hinges on an effective representation of sketch images and an efficient search method. In this paper, we propose a representation scheme which takes sketch strokes into account with local features, thereby facilitating efficient retrieval with codebooks. Stroke features are detected via densely sampled points on stroke lines from which local gradients are further enhanced and described by a quantized histogram of gradients. A codebook is organized in a hierarchical vocabulary tree, which maintains structural information of visual words and enables efficient retrieval in sub-linear time. Experimental results on three data sets demonstrate the merits of the proposed algorithm for effective and efficient sketch retrieval.