In Semi-supervised learning(SSL), we always accept cluster assumption, assuming features in different high-density regions belong to other categories. However, it is always ignored by existing algorithms and needs mathematical explanations. This paper first proposes a theorem to statistically explain cluster assumption and prove that the probability density can significantly help to use the prior fully. A Probability-Density-Aware Measure(PM) is proposed based on the theorem to discern the similarity between neighbor points. The PM is deployed to improve Label Propagation and a new pseudo-labeling algorithm, the Probability-Density-Aware Label Propagation(PMLP), is proposed. We also prove that traditional first-order similarity pseudo-labeling could be viewed as a particular case of PMLP, which provides a comprehensive theoretical understanding of PMLP's superior performance. Extensive experiments demonstrate that PMLP achieves outstanding performance compared with other recent methods.

Knowledge distillation (KD) is a valuable yet challenging approach that enhances a compact student network by learning from a high-performance but cumbersome teacher model. However, previous KD methods for image restoration overlook the state of the student during the distillation, adopting a fixed solution space that limits the capability of KD. Additionally, relying solely on L1-type loss struggles to leverage the distribution information of images. In this work, we propose a novel dynamic contrastive knowledge distillation (DCKD) framework for image restoration. Specifically, we introduce dynamic contrastive regularization to perceive the student's learning state and dynamically adjust the distilled solution space using contrastive learning. Additionally, we also propose a distribution mapping module to extract and align the pixel-level category distribution of the teacher and student models. Note that the proposed DCKD is a structure-agnostic distillation framework, which can adapt to different backbones and can be combined with methods that optimize upper-bound constraints to further enhance model performance. Extensive experiments demonstrate that DCKD significantly outperforms the state-of-the-art KD methods across various image restoration tasks and backbones.

In the quest for artificial general intelligence, Multi-modal Large Language Models (MLLMs) have emerged as a focal point in recent advancements. However, the predominant focus remains on developing their capabilities in static image understanding. The potential of MLLMs to process sequential visual data is still insufficiently explored, highlighting the lack of a comprehensive, high-quality assessment of their performance. In this paper, we introduce Video-MME, the first-ever full-spectrum, Multi-Modal Evaluation benchmark of MLLMs in Video analysis. Our work distinguishes from existing benchmarks through four key features: 1) Diversity in video types, spanning 6 primary visual domains with 30 subfields to ensure broad scenario generalizability; 2) Duration in temporal dimension, encompassing both short-, medium-, and long-term videos, ranging from 11 seconds to 1 hour, for robust contextual dynamics; 3) Breadth in data modalities, integrating multi-modal inputs besides video frames, including subtitles and audios, to unveil the all-round capabilities of MLLMs; 4) Quality in annotations, utilizing rigorous manual labeling by expert annotators to facilitate precise and reliable model assessment. With Video-MME, we extensively evaluate various state-of-the-art MLLMs, and reveal that Gemini 1.5 Pro is the best-performing commercial model, significantly outperforming the open-source models with an average accuracy of 75%, compared to 71.9% for GPT-4o. The results also demonstrate that Video-MME is a universal benchmark that applies to both image and video MLLMs. Further analysis indicates that subtitle and audio information could significantly enhance video understanding. Besides, a decline in MLLM performance is observed as video duration increases for all models. Our dataset along with these findings underscores the need for further improvements in handling longer sequences and multi-modal data, shedding light on future MLLM development. Project page: https://video-mme.github.io.

Deep learning has significantly advanced the object detection field. However, tiny object detection (TOD) remains a challenging problem. We provide a new analysis method to examine the TOD challenge through occlusion-based attribution analysis in the frequency domain. We observe that tiny objects become less distinct after feature encoding and can benefit from the removal of high-frequency information. In this paper, we propose a novel approach named Spectral Enhancement for Tiny object detection (SET), which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate the effectiveness of our method. Especially, SET boosts the prior art RFLA by 3.2% AP on the AI-TOD dataset.

Weakly supervised semantic segmentation (WSSS) has garnered considerable attention due to its effective reduction of annotation costs. Most approaches utilize Class Activation Maps (CAM) to produce pseudo-labels, thereby localizing target regions using only image-level annotations. However, the prevalent methods relying on vision transformers (ViT) encounter an "over-expansion" issue, i.e., CAM incorrectly expands high activation value from the target object to the background regions, as it is difficult to learn pixel-level local intrinsic inductive bias in ViT from weak supervisions. To solve this problem, we propose a Progressive Confidence Region Expansion (PCRE) framework for WSSS, it gradually learns a faithful mask over the target region and utilizes this mask to correct the confusion in CAM. PCRE has two key components: Confidence Region Mask Expansion (CRME) and Class-Prototype Enhancement (CPE). CRME progressively expands the mask in the small region with the highest confidence, eventually encompassing the entire target, thereby avoiding unintended coverage of background areas. CPE aims to enhance mask generation in CRME by leveraging the similarity between the learned, dataset-level class prototypes and patch features as supervision to optimize the mask output from CRME. Extensive experiments demonstrate that our method outperforms the existing single-stage and multi-stage approaches on the PASCAL VOC and MS COCO benchmark. Our code is available at https://github.com/xxf011/WSSS-PCRE.

Current weakly supervised point cloud semantic segmentation struggles with insufficient utilization of limited annotations in unimodal representation learning due to the sparse and textureless nature of point clouds. In this work, we leverage cross-modality information by transferring knowledge from image and text sources to the point cloud network. The intuition is that images contribute rich texture, color, and discriminative information, complementing point clouds to boost semantic segmentation performance. To reduce extensive computational resources for cross-modality fusion, we introduce the Multi-Scale Deformable Knowledge Transfer, an innovative training scheme that optimizes and extends the one-to-one mapping to flexible one-to-many relations between multi-modal data. Furthermore, we employ pre-trained image-text models to generate pseudo labels for point clouds and construct positive and negative samples for semantic contrastive regularization, facilitating the full exploitation of unlabeled data. The experimental results evaluated on SemanticKITTI and nuScenes demonstrate substantial improvements, achieving an average gain of 3.8% over the previous weakly supervised methods, and comparable performances to fully supervised approaches.

Leveraging the complementary characteristics of visible (RGB) and infrared (IR) imagery offers significant potential for improving object detection. In this paper, we propose WaveMamba, a cross-modality fusion method that efficiently integrates the unique and complementary frequency features of RGB and IR decomposed by Discrete Wavelet Transform (DWT). An improved detection head incorporating the Inverse Discrete Wavelet Transform (IDWT) is also proposed to reduce information loss and produce the final detection results. The core of our approach is the introduction of WaveMamba Fusion Block (WMFB), which facilitates comprehensive fusion across low-/high-frequency sub-bands. Within WMFB, the Low-frequency Mamba Fusion Block (LMFB), built upon the Mamba framework, first performs initial low-frequency feature fusion with channel swapping, followed by deep fusion with an advanced gated attention mechanism for enhanced integration. High-frequency features are enhanced using a strategy that applies an ``absolute maximum" fusion approach. These advancements lead to significant performance gains, with our method surpassing state-of-the-art approaches and achieving average mAP improvements of 4.5% on four benchmarks.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in vision understanding, reasoning, and interaction. However, the inference computation and memory increase progressively with the generation of output tokens during decoding, directly affecting the efficacy of MLLMs. Existing methods attempt to reduce the vision context redundancy to achieve efficient MLLMs. Unfortunately, the efficiency benefits of the vision context reduction in the prefill stage gradually diminish during the decoding stage. To address this problem, we proposed a dynamic vision-language context sparsification framework Dynamic-LLaVA, which dynamically reduces the redundancy of vision context in the prefill stage and decreases the memory and computation overhead of the generated language context during decoding. Dynamic-LLaVA designs a tailored sparsification inference scheme for different inference modes, i.e., prefill, decoding with and without KV cache, to achieve efficient inference of MLLMs. In practice, Dynamic-LLaVA can reduce computation consumption by $\sim$75\% in the prefill stage. Meanwhile, throughout the entire generation process of MLLMs, Dynamic-LLaVA reduces the $\sim$50\% computation consumption under decoding without KV cache, while saving $\sim$50\% GPU memory overhead when decoding with KV cache, due to the vision-language context sparsification. Extensive experiments also demonstrate that Dynamic-LLaVA achieves efficient inference for MLLMs with negligible understanding and generation ability degradation or even performance gains compared to the full-context inference baselines. Code is available at https://github.com/Osilly/dynamic_llava.

Knowledge distillation (KD) is a promising yet challenging model compression approach that transmits rich learning representations from robust but resource-demanding teacher models to efficient student models. Previous methods for image super-resolution (SR) are often tailored to specific teacher-student architectures, limiting their potential for improvement and hindering broader applications. This work presents a novel KD framework for SR models, the multi-granularity Mixture of Priors Knowledge Distillation (MiPKD), which can be universally applied to a wide range of architectures at both feature and block levels. The teacher’s knowledge is effectively integrated with the student's feature via the Feature Prior Mixer, and the reconstructed feature propagates dynamically in the training phase with the Block Prior Mixer. Extensive experiments illustrate the significance of the proposed MiPKD technique.

Knowledge distillation (KD) compresses deep neural networks by transferring task-related knowledge from cumbersome pre-trained teacher models to more compact student models. However, vanilla KD for image super-resolution (SR) networks yields only limited improvements due to the inherent nature of SR tasks, where the outputs of teacher models are noisy approximations of high-quality label images. In this work, we show that the potential of vanilla KD has been underestimated and demonstrate that the ingenious application of data augmentation methods can close the gap between it and more complex, well-designed methods. Unlike conventional training processes typically applying image augmentations simultaneously to both low-quality inputs and high-quality labels, we propose AugKD utilizing unpaired data augmentations to 1) generate auxiliary distillation samples and 2) impose label consistency regularization. Comprehensive experiments show that the AugKD significantly outperforms existing state-of-the-art KD methods across a range of SR tasks.

Open-vocabulary object detection seeks to recognize objects from arbitrary language inputs, extending detection beyond fixed training categories. While recent methods have made progress in detecting unseen categories, they typically require a set of predefined categories during the inference stage, hindering practical deployment in open-world scenarios. To overcome this crucial limitation, we propose UniPerception , a novel universal perception framework based on open-vocabulary object detection. It not only excels at open-vocabulary object detection but is also capable of generating labels for target objects in the absence of predefined vocabularies, and can be adapted to a broad range of vision-language tasks simply by modifying the language instructions. UniPerception seamlessly integrates three key innovations: 1) a robust visual detector trained on diverse data sources to capture rich and generalizable visual representations; 2) a language model with interleaved cross-modality fusion layers to interpret instructions and generate fine-grained responses conditioned on visual features; and 3) a tailored multi-stage training strategy that effectively bridges detection-specific learning with general vision-language understanding. We conduct extensive experiments on multiple benchmarks for open-vocabulary object detection (COCO, LVIS, ODinW), referring expression comprehension (RefCOCO/+/g, D3), and vision-language understanding (Flickr30k, VQAv2, GQA). The results show that UniPerception achieves strong open-world generalization and multi-modal understanding, outperforming the existing state-of-the-art methods and establishing itself as a unified, instruction-driven perception system.

Soccer is a globally popular sporting event, typically characterized by long matches and distinctive highlight moments. Recent advances in Multimodal Large Language Models (MLLMs) show promising capabilities in temporal grounding and video understanding. However, generating soccer commentary requires both precise temporal localization and semantically rich descriptions over long-form videos. Existing soccer MLLMs often rely on temporal priors for caption generation, which limits their ability to process the entire video in an end-to-end manner. Traditional approaches, on the other hand, follow a complex two-step paradigm that fails to capture the global context, leading to suboptimal performance. To solve the above issues, we present TimeSoccer, the first end-to-end soccer MLLM for Single-anchor Dense Video Captioning (SDVC) in full-match soccer videos. TimeSoccer jointly predicts timestamps and generates captions in a single pass, enabling global context modeling across 45-minute matches. To support long video understanding of soccer matches, we introduce MoFA-Select, a training-free, motion-aware frame compression module that adaptively selects representative frames via a coarse-to-fine strategy, and incorporates complementary training paradigms to strengthen the model's ability to handle long temporal sequences. Extensive experiments demonstrate that our TimeSoccer achieves State-of-The-Art (SoTA) performance on the SDVC task in an end-to-end form, generating high-quality commentary with accurate temporal alignment and strong semantic relevance. For more information, please visit: https://vpx-ecnu.github.io/TimeSoccer-Website/.

As demand for virtual digital characters grows in fields such as virtual reality, gaming, and animation, generating highly controllable human motion within scenes has become a key research focus. Existing methods for scene-aware motion generation typically rely on global alignment or latent space matching, which provides limited control over the fine-grained movements of individual body parts. This limitation often leads to rigid and unrealistic motions when interacting with complex environments. Therefore, we propose the Body-Aware Interaction Diffusion Model (BA-IDM), which enables fine-grained control of human motion within a scene by leveraging multimodal information. Text descriptions, motion scenes, and movement trajectories can all serve as inputs, allowing for precise control of each body part and facilitating the generation of a wide range of complex actions. Moreover, our approach is designed to operate on de-identified motion data, effectively protecting user privacy throughout the process, which is essential for practical and user-centric applications.

Recent advances in Multimodal Large Language Models (MLLMs) have significantly improved 2D visual understanding, prompting interest in their application to complex 3D reasoning tasks. However, it remains unclear whether these models can effectively capture the detailed spatial information required for robust real-world performance, especially cross-view consistency, a key requirement for accurate 3D reasoning. Considering this issue, we introduce Viewpoint Learning, a task designed to evaluate and improve the spatial reasoning capabilities of MLLMs. We present the Viewpoint-100K dataset, consisting of 100K object-centric image pairs with diverse viewpoints and corresponding question-answer pairs. Our approach employs a two-stage fine-tuning strategy: first, foundational knowledge is injected to the baseline MLLM via Supervised Fine-Tuning (SFT) on Viewpoint-100K, resulting in significant improvements across multiple tasks; second, generalization is enhanced through Reinforcement Learning using the Group Relative Policy Optimization (GRPO) algorithm on a broader set of questions. Additionally, we introduce a hybrid cold-start initialization method designed to simultaneously learn viewpoint representations and maintain coherent reasoning thinking. Experimental results show that our approach significantly activates the spatial reasoning ability of MLLM, improving performance on both in-domain and out-of-domain reasoning tasks. Our findings highlight the value of developing foundational spatial skills in MLLMs, supporting future progress in robotics, autonomous systems, and 3D scene understanding.

Despite Transformers have achieved significant success in low-level vision tasks, they are constrained by computing self-attention with a quadratic complexity and limited-size windows. This limitation results in a lack of global receptive field across the entire image. Recently, State Space Models (SSMs) have gained widespread attention due to their global receptive field and linear complexity with respect to input length. However, integrating SSMs into low-level vision tasks presents two major challenges: 1) Relationship degradation of long-range tokens with a long-range forgetting problem by encoding pixel-by-pixel high-resolution images. 2) Significant redundancy in the existing multi-direction scanning strategy. To this end, we propose Hi-Mamba for image super-resolution (SR) to address these challenges, which unfolds the image with only a single scan. Specifically, the Global Hierarchical Mamba Block (GHMB) enables token interactions across the entire image, providing a global receptive field while leveraging a multi-scale structure to facilitate long-range dependency learning. Additionally, the Direction Alternation Module (DAM) adjusts the scanning patterns of GHMB across different layers to enhance spatial relationship modeling. Extensive experiments demonstrate that our Hi-Mamba achieves 0.2-0.27dB PSNR gains on the Urban100 dataset across different scaling factors compared to the state-of-the-art MambaIRv2 for SR. Moreover, our lightweight Hi-Mamba also outperforms lightweight SRFormer by 0.39dB PSNR for $\times 2$ SR.

Transformer is leading a trend in the field of image processing. While existing lightweight image processing transformers have achieved notable success, they primarily focus on reducing FLOPs (floating-point operations) or the number of parameters, rather than on practical inference acceleration. In this paper, we present a latency-aware image processing transformer, termed LIPT. We devise the low-latency proportion LIPT block that substitutes memory-intensive operators with the combination of self-attention and convolutions to achieve practical speedup. Specifically, we propose a novel non-volatile sparse masking self-attention (NVSM-SA) that utilizes a pre-computing sparse mask to capture contextual information from a larger window with no extra computation overload. Besides, a high-frequency reparameterization module (HRM) is proposed to make LIPT block reparameterization friendly, enhancing the model's ability to reconstruct fine details. Extensive experiments on multiple image processing tasks (e.g., image super-resolution (SR), JPEG artifact reduction, and image denoising) demonstrate the superiority of LIPT on both latency and PSNR. LIPT achieves real-time GPU inference with state-of-the-art performance on multiple image SR benchmarks. The source codes are released at https://github.com/Lucien66/LIPT.

Cross-modality object detection aims to fuse complementary information from different modalities to improve model performance, which achieves a wider range of applications. However, traditional cross-modality fusion methods, based on CNN or Transformer, inadequately address the issue of pseudo-target information, which causes model attention dispersion to degrade object detection performance. In this paper, we investigate a novel cross-modality fusion approach by associating cross-modal features in a hidden state space based on an improved Mamba with a gating attention mechanism. We propose theFusion-Mamba Block(FMB), designed to map cross-modal features into a hidden state space for interaction, thereby refining the model’s attention on true target areas and enhancing overall performance. The FMB comprises two key modules: State Space Channel Swapping (SSCS) module, which facilitates the fusion of shallow features, and Dual State Space Fusion (DSSF) module, which enables deep fusion and effectively suppresses pseudo-target information within the hidden state space. Our proposed method outperforms state-of-the-art approaches, achieving improvements of 5.9%, 3.5% and 2.1% mAP on$M^{3}$FD, DroneVehicle and FLIR-Aligned, respectively. To the best of our knowledge, this work establishes a new baseline for cross-modality object detection, providing a robust foundation for future research in this area.

Graph neural networks (GNNs) has demonstrated its capabilities in the field of scene graph generation (SGG) by updating node representations from neighboring nodes. Actually it can be viewed as a form of low-pass filter in the spatial domain, which smooths node feature representation and retains commonalities among nodes. However, spatial GNNs does not work well in the case of heterophilic SGG in which fine-grained predicates are always connected to a large number of coarse-grained predicates. Blind smoothing undermines the discriminative information of the fine-grained predicates, resulting in failure to predict them accurately. To address the heterophily, our key idea is to design tailored filters by wavelet transform from the spectral domain. First, we prove rigorously that when the heterophily on the scene graph increases, the spectral energy gradually shifts towards the high-frequency part. Inspired by this observation, we subsequently propose the Kumaraswamy Wavelet Graph Neural Network (KWGNN). KWGNN leverages complementary multi-group Kumaraswamy wavelets to cover all frequency bands. Finally, KWGNN adaptively generates band-pass filters and then integrates the filtering results to better accommodate varying levels of smoothness on the graph. Comprehensive experiments on the Visual Genome and Open Images datasets show that our method achieves state-of-the-art performance.

Symmetric positive definite(SPD) matrices have shown important value and applications in statistics and machine learning, such as FMRI analysis and traffic prediction. Previous works on SPD matrices mostly focus on discriminative models, where predictions are made directly on E(X|y), where y is a vector and X is an SPD matrix. However, these methods are challenging to handle for large-scale data. In this paper, inspired by denoising diffusion probabilistic model(DDPM), we propose a novel generative model, termed SPD-DDPM, by introducing Gaussian distribution in the SPD space to estimate E(X|y). Moreover, our model can estimate p(X) unconditionally and flexibly without giving y. On the one hand, the model conditionally learns p(X|y) and utilizes the mean of samples to obtain E(X|y) as a prediction. On the other hand, the model unconditionally learns the probability distribution of the data p(X) and generates samples that conform to this distribution. Furthermore, we propose a new SPD net which is much deeper than the previous networks and allows for the inclusion of conditional factors. Experiment results on toy data and real taxi data demonstrate that our models effectively fit the data distribution both unconditionally and conditionally.

Despite weakly supervised object detection (WSOD) being a promising step toward evading strong instance-level annotations, its capability is confined to closed-set categories within a single training dataset. In this paper, we propose a novel weakly supervised open-vocabulary object detection framework, namely WSOVOD, to extend traditional WSOD to detect novel concepts and utilize diverse datasets with only image-level annotations. To achieve this, we explore three vital strategies, including dataset-level feature adaptation, image-level salient object localization, and region-level vision-language alignment. First, we perform data-aware feature extraction to produce an input-conditional coefficient, which is leveraged into dataset attribute prototypes to identify dataset bias and help achieve cross-dataset generalization. Second, a customized location-oriented weakly supervised region proposal network is proposed to utilize high-level semantic layouts from the category-agnostic segment anything model to distinguish object boundaries. Lastly, we introduce a proposal-concept synchronized multiple-instance network, i.e., object mining and refinement with visual-semantic alignment, to discover objects matched to the text embeddings of concepts. Extensive experiments on Pascal VOC and MS COCO demonstrate that the proposed WSOVOD achieves new state-of-the-art compared with previous WSOD methods in both close-set object localization and detection tasks. Meanwhile, WSOVOD enables cross-dataset and open-vocabulary learning to achieve on-par or even better performance than well-established fully-supervised open-vocabulary object detection (FSOVOD).

DEtection TRansformer (DETR)-based models have achieved remarkable performance. However, they are accompanied by a large computation overhead cost, which significantly prevents their applications on resource-limited devices. Prior arts attempt to reduce the computational burden of DETR using low-bit quantization, while these methods sacrifice a severe significant performance on weight-activation-attention low-bit quantization. We observe that the number of matching queries and positive samples affect much on the representation capacity of queries in DETR, while quantifying queries of DETR further reduces its representational capacity, thus leading to a severe performance drop. We introduce a new quantization strategy based on Auxiliary Queries for DETR (AQ-DETR), aiming to enhance the capacity of quantized queries. In addition, a layer-by-layer distillation is proposed to reduce the quantization error between quantized attention and full-precision counterpart. Through our extensive experiments on large-scale open datasets, the performance of the 4-bit quantization of DETR and Deformable DETR models is comparable to full-precision counterparts.

3D Scene Graph Generation (3DSGG) aims to classify objects and their predicates within 3D point cloud scenes. However, current 3DSGG methods struggle with two main challenges. 1) The dependency on labor-intensive ground-truth annotations. 2) Closed-set classes training hampers the recognition of novel objects and predicates. Addressing these issues, our idea is to extract cross-modality features by CLIP from text and image data naturally related to 3D point clouds. Cross-modality features are used to train a robust 3D scene graph (3DSG)feature extractor. Specifically, we propose a novel Cross-Modality Contrastive Learning 3DSGG (CCL-3DSGG) method. Firstly, to align the text with 3DSG, the text is parsed into word level that are consistent with the 3DSG annotation. To enhance robustness during the alignment, adjectives are exchanged for different objects as negative samples. Then, to align the image with 3DSG, the camera view is treated as a positive sample and other views as negatives. Lastly, the recognition of novel object and predicate classes is achieved by calculating the cosine similarity between prompts and 3DSG features. Our rigorous experiments confirm the superior open-vocabulary capability and applicability of CCL-3DSGG in real-world contexts.

The remarkable performance of Vision Transformers (ViTs) typically requires an extremely large training cost. Existing methods have attempted to accelerate the training of ViTs, yet typically disregard method universality with accuracy dropping. Meanwhile, they break the training consistency of the original transformers, including the consistency of hyperparameters, architecture, and strategy, which prevents them from being widely applied to different Transformer networks. In this paper, we propose a novel token growth scheme Token Expansion (termed ToE) to achieve consistent training acceleration for ViTs. We introduce an “initialization-expansion-merging” pipeline to maintain the integrity of the intermediate feature distribution of original transformers, preventing the loss of crucial learnable information in the training process. ToE can not only be seamlessly integrated into the training and fine-tuning process of transformers (e.g., DeiT and LV-ViT), but also effective for efficient training frameworks (e.g., EfficientTrain), without twisting the original training hyperparameters, architecture, and introducing additional training strategies. Extensive experiments demonstrate that ToE achieves about 1.3× faster for the training of ViTs in a lossless manner, or even with performance gains over the full-token training baselines. Code is available at https://github.com/Osilly/TokenExpansion.

Vision foundation models have been explored recently to build general-purpose vision systems. However, predomi-nant paradigms, driven by casting instance-level tasks as an object-word alignment, bring heavy cross-modality in-teraction, which is not effective in prompting object detection and visual grounding. Another line of work that fo-cuses on pixel-level tasks often encounters a large annotation gap of things and stuff, and suffers from mutual inter-ference between foreground-object and background-class segmentation. In stark contrast to the prevailing methods, we present APE, a universal visual perception model for aligning and prompting everything all at once in an image to perform diverse tasks, i.e., detection, segmentation, and grounding, as an instance-level sentence-object matching paradigm. Specifically, APE advances the convergence of detection and grounding by reformulating language-guided grounding as open-vocabulary detection, which efficiently scales up model prompting to thousands of category vocab-ularies and region descriptions while maintaining the ef-fectiveness of cross-modality fusion. To bridge the granu-larity gap of different pixel-level tasks, APE equalizes se-mantic and panoptic segmentation to proxy instance learning by considering any isolated regions as individual in-stances. APE aligns vision and language representation on broad data with natural and challenging characteristics all at once without task-specific fine-tuning. The extensive ex-periments on over 160 datasets demonstrate that, with only one-suit of weights, APE outperforms (or is on par with) the state-of-the-art models, proving that an effective yet univer-sal perception for anything aligning and prompting is in-deed feasible. Codes and trained models are released at https://github.com/shenyunhang/APE.

The CLIP network excels in various tasks, but struggles with text-visual images i.e., images that contain both text and visual objects; it risks confusing textual and visual representations. To address this issue, we propose MirrorCLIP, a zero-shot framework, which disentangles the image features of CLIP by exploiting the difference in the mirror effect between visual objects and text in the images. Specifically, MirrorCLIP takes both original and flipped images as inputs, comparing their features dimension-wise in the latent space to generate disentangling masks. With disentangling masks, we further design filters to separate textual and visual factors more precisely, and then get disentangled representations. Qualitative experiments using stable diffusion models and class activation mapping (CAM) validate the effectiveness of our disentanglement. Moreover, our proposed MirrorCLIP reduces confusion when encountering text-visual images and achieves a substantial improvement on typographic defense, further demonstrating its superior ability of disentanglement. Our code is available at https://github.com/tcwangbuaa/MirrorCLIP

Semi-supervised learning (SSL), thanks to the significant reduction of data annotation costs, has been an active research topic for large-scale 3D scene understanding. However, the existing SSL-based methods suffer from severe training bias, mainly due to class imbalance and long-tail distributions of the point cloud data. As a result, they lead to a biased prediction for the tail class segmentation. In this paper, we introduce a new decoupling optimization framework, which disentangles feature representation learning and classifier in an alternative optimization manner to shift the bias decision boundary effectively. In particular, we first employ two-round pseudo-label generation to select unlabeled points across head-to-tail classes. We further introduce multi-class imbalanced focus loss to adaptively pay more attention to feature learning across head-to-tail classes. We fix the backbone parameters after feature learning and retrain the classifier using ground-truth points to update its parameters. Extensive experiments demonstrate the effectiveness of our method outperforming previous state-of-the-art methods on both indoor and outdoor 3D point cloud datasets ( i.e. , S3DIS, ScanNet-V2, Semantic3D, and SemanticKITTI) using 1% and 1pt evaluation.

Deep learning (DL) powered real-time applications usually need continuous training using data streams generated over time and across different geographical locations. Enabling data offloading among computation nodes through model training is promising to mitigate the problem that devices generating large datasets may have low computation capability. However, offloading can compromise model convergence and incur communication costs, which must be balanced with the long-term cost spent on computation and model synchronization. Therefore, this paper proposes EdgeC3, a novel framework that can optimize the frequency of model aggregation and dynamic offloading for continuously generated data streams, navigating the trade-off between long-term accuracy and cost. We first provide a new error bound to capture the impacts of data dynamics that are varying over time and heterogeneous across devices, as well as quantifying varied data heterogeneity between local models and the global one. Based on the bound, we design a two-timescale online optimization framework. We periodically learn the synchronization frequency to adapt with uncertain future offloading and network changes. In the finer timescale, we manage online offloading by extending Lyapunov optimization techniques to handle an unconventional setting, where our long-term global constraint can have abruptly changed aggregation frequencies that are decided in the longer timescale. Finally, we theoretically prove the convergence of EdgeC3 by integrating the coupled effects of our two-timescale decisions, and we demonstrate its advantage through extensive experiments performing distributed DL training for different domains.

Cross-domain crowd counting has shown progressively improved performance. However, most methods fail to explicitly consider the transferability of different features between source and target domains. In this paper, we propose an innovative explicit Invariant Feature induced Cross-domain Knowledge Transformation framework to address the inconsistent domain-invariant features of different domains. The main idea is to explicitly extract domain-invariant features from both source and target domains, which builds a bridge to transfer more rich knowledge between two domains. The framework consists of three parts, global feature decoupling (GFD), relation exploration and alignment (REA), and graph-guided knowledge enhancement (GKE). In the GFD module, domain-invariant features are efficiently decoupled from domain-specific ones in two domains, which allows the model to distinguish crowds features from backgrounds in the complex scenes. In the REA module both inter-domain relation graph (Inter-RG) and intra-domain relation graph (Intra-RG) are built. Specifically, Inter-RG aggregates multi-scale domain-invariant features between two domains and further aligns local-level invariant features. Intra-RG preserves taskrelated specific information to assist the domain alignment. Furthermore, GKE strategy models the confidence of pseudolabels to further enhance the adaptability of the target domain. Various experiments show our method achieves state-of-theart performance on the standard benchmarks. Code is available at https://github.com/caiyiqing/IF-CKT.

Online Knowledge Distillation (OKD) is designed to alleviate the dilemma that the high-capacity pre-trained teacher model is not available. However, the existing methods mostly focus on improving the ensemble prediction accuracy from multiple students (a.k.a. branches), which often overlook the homogenization problem that makes student model saturate quickly and hurts the performance. We assume that the intrinsic bottleneck of the homogenization problem comes from the identical branch architecture and coarse ensemble strategy. We propose a novel Adaptive Hierarchy-Branch Fusion framework for Online Knowledge Distillation, termed AHBF-OKD, which designs hierarchical branches and adaptive hierarchy-branch fusion module to boost the model diversity and aggregate complementary knowledge. Specifically, we first introduce hierarchical branch architectures to construct diverse peers by increasing the depth of branches monotonously on the basis of target branch. To effectively transfer knowledge from the most complex branch to the simplest target branch, we propose an adaptive hierarchy-branch fusion module to create hierarchical teacher assistants recursively, which regards the target branch as the smallest teacher assistant. During the training, the teacher assistant from the previous hierarchy is explicitly distilled by the teacher assistant and the branch from the current hierarchy. Thus, the important scores to different branches are effectively and adaptively allocated to reduce the branch homogenization. Extensive experiments demonstrate the effectiveness of AHBF-OKD on different datasets, including CIFAR-10/100 and ImageNet 2012. For example, on ImageNet 2012, the distilled ResNet-18 achieves Top-1 error of 29.28\%, which significantly outperforms the state-of-the-art methods. The source code is available at https://github.com/linruigong965/AHBF.

No abstract available.

Continual learning aims to enable a model to incrementally learn knowledge from sequentially arrived data. Previous works adopt the conventional classification architecture, which consists of a feature extractor and a classifier. The feature extractor is shared across sequentially arrived tasks or classes, but one specific group of weights of the classifier corresponding to one new class should be incrementally expanded. Consequently, the parameters of a continual learner gradually increase. Moreover, as the classifier contains all historical arrived classes, a certain size of the memory is usually required to store rehearsal data to mitigate classifier bias and catastrophic forgetting. In this paper, we propose a non-incremental learner, named AttriCLIP, to incrementally extract knowledge of new classes or tasks. Specifically, AttriCLIP is built upon the pre-trained visual-language model CLIP. Its image encoder and text encoder are fixed to extract features from both images and text. Text consists of a category name and a fixed number of learnable parameters which are selected from our designed attribute word bank and serve as attributes. As we compute the visual and textual similarity for classification, AttriCLIP is a non-incremental learner. The attribute prompts, which encode the common knowledge useful for classification, can effectively mitigate the catastrophic forgetting and avoid constructing a replay memory. We evaluate our AttriCLIP and compare it with CLIP-based and previous state-of-the-art continual learning methods in realistic settings with domain-shift and long-sequence learning. The results show that our method performs favorably against previous state-of-the-arts. The implementation code will be available at https://gitee.com/mindspore/models/tree/master/research/cv/AttriCLIP.

Brain tumor anomaly detection plays a critical role in the field of computer-aided diagnosis, which has attracted ever-increasing focus from the medical community However, brain tumor data are scarce and tough to classify. Unsupervised methods enable the reduction of huge labeling costs to be applied to brain tumor anomaly detection during the training only given normal brain images. However, the existing unsupervised methods distinguish whether the input image is abnormal in the image space, which cannot effectively learn the discriminative features. In this paper, we propose a novel brain tumor anomaly detection method via Latent Feature Regularization based Adversarial Network (LFRA-Net), which leverages a latent feature regularizer into adversarial learning to obtain the discriminative features. Comprehensive experiments on BraTS, HCP, MNIST, and CIFAR-10 datasets evaluate the effectiveness of our LFRANet, which outperforms state-of-the-art unsupervised learning methods.

The task of dynamic scene graph generation (DSGG) aims at constructing a set of frame-level scene graphs for the given video. It suffers from two kinds of spurious correlation problems. First, the spurious correlation between input object pair and predicate label is caused by the biased predicate sample distribution in dataset. Second, the spurious correlation between contextual information and predicate label arises from interference caused by background content in both the current frame and adjacent frames of the video sequence. To alleviate spurious correlations, our work is formulated into two sub-tasks: video-specific commonsense graph generation (VsCG) and causal inference (CI). VsCG module aims to alleviate the first correlation by integrating prior knowledge into prediction. Information of all the frames in current video is used to enhance the commonsense graph constructed from co-occurrence patterns of all training samples. Thus, the commonsense graph has been augmented with video-specific temporal dependencies. Then, a CI strategy with both intervention and counterfactual is used. The intervention component further eliminates the first correlation by forcing the model to consider all possible predicate categories fairly, while the counterfactual component resolves the second correlation by removing the bad effect from context. Comprehensive experiments on the Action Genome dataset show that the proposed method achieves state-of-the-art performance.

Deep learning (DL) powered real-time applications usually need continuous training using data streams generated geographically. Enabling data offloading among computation nodes through model training is promising to mitigate the problem that devices generating large datasets may have low computation capability. However, offloading can compromise model convergence and incur communication costs, which must be balanced with the cost spent on computation and model synchronization. Therefore, this paper proposes EdgeC3, a novel framework that can optimize the frequency of model aggregation and dynamic offloading for continuously generated data streams, navigating the trade-off between long-term accuracy and cost. We first provide a new error bound to capture the impacts of data dynamics that are varying over time and heterogeneous across devices. Based on the bound, we design a two-timescale online optimization framework. We periodically learn the synchronization frequency to adapt with uncertain future offloading and network changes. In the finer timescale, we manage online offloading by extending Lyapunov optimization techniques to handle an unconventional setting, where our long-term global constraint can have abruptly changed aggregation frequencies that are decided in the longer timescale. Finally, we theoretically prove the convergence of EdgeC3 by integrating the coupled effects of our two-timescale decisions, and we demonstrate its advantage through extensive experiments.

U-Nets have achieved tremendous success in medical image segmentation. Nevertheless, it may have limitations in global (long-range) contextual interactions and edge-detail preservation. In contrast, the Transformer module has an excellent ability to capture long-range dependencies by leveraging the self-attention mechanism into the encoder. Although the Transformer module was born to model the long-range dependency on the extracted feature maps, it still suffers high computational and spatial complexities in processing high-resolution 3D feature maps. This motivates us to design an efficient Transformer-based UNet model and study the feasibility of Transformer-based network architectures for medical image segmentation tasks. To this end, we propose to self-distill a Transformer-based UNet for medical image segmentation, which simultaneously learns global semantic information and local spatial-detailed features. Meanwhile, a local multi-scale fusion block is first proposed to refine fine-grained details from the skipped connections in the encoder by the main CNN stem through self-distillation, only computed during training and removed at inference with minimal overhead. Extensive experiments on BraTS 2019 and CHAOS datasets show that our MISSU achieves the best performance over previous state-of-the-art methods. Code and models are available at: https://github.com/wangn123/MISSU.git.

Video anomaly detection aims to automatically identify unusual objects or behaviours by learning from normal videos. Previous methods tend to use simplistic reconstruction or prediction constraints, which leads to the insufficiency of learned representations for normal data. As such, we propose a novel bi-directional architecture with three consistency constraints to comprehensively regularize the prediction task from pixel-wise, cross-modal, and temporal-sequence levels. First, predictive consistency is proposed to consider the symmetry property of motion and appearance in forwards and backwards time, which ensures the highly realistic appearance and motion predictions at the pixel-wise level. Second, association consistency considers the relevance between different modalities and uses one modality to regularize the prediction of another one. Finally, temporal consistency utilizes the relationship of the video sequence and ensures that the predictive network generates temporally consistent frames. During inference, the pattern of abnormal frames is unpredictable and will therefore cause higher prediction errors. Experiments show that our method outperforms advanced anomaly detectors and achieves state-of-the-art results on UCSD Ped2, CUHK Avenue, and ShanghaiTech datasets.

To address the huge labeling cost in large-scale point cloud semantic segmentation, we propose a novel hybrid contrastive regularization (HybridCR) framework in weakly-supervised setting, which obtains competitive performance compared to its fully-supervised counterpart. Specifically, HybridCR is the first framework to leverage both point consistency and employ contrastive regularization with pseudo labeling in an end-to-end manner. Fundamentally, HybridCR explicitly and effectively considers the semantic similarity between local neighboring points and global characteristics of 3D classes. We further design a dynamic point cloud augmentor to generate diversity and robust sample views, whose transformation parameter is jointly optimized with model training. Through extensive experiments, HybridCR achieves significant performance improvement against the SOTA methods on both indoor and outdoor datasets, e.g., S3DIS, ScanNet-V2, Semantic3D, and SemanticKITTI.

Transformers have shown remarkable progress on computer vision tasks in the past year. Compared to their CNN counterparts, transformers usually need the help of distillation to achieve comparable results on middle or small sized datasets. Meanwhile, recent researches discover that when transformers are trained with supervised and self-supervised manner respectively, the captured patterns are quite different both qualitatively and quantitatively. These findings motivate us to introduce an self-supervised teaching assistant (SSTA) besides the commonly used supervised teacher to improve the performance of transformers. Specifically, we propose a head-level knowledge distillation method that selects the most important head of the supervised teacher and self-supervised teaching assistant, and let the student mimic the attention distribution of these two heads, so as to make the student focus on the relationship between tokens deemed by the teacher and the teacher assistant. Extensive experiments verify the effectiveness of SSTA and demonstrate that the proposed SSTA is a good compensation to the supervised teacher. Meanwhile, some analytical experiments towards multiple perspectives (e.g. prediction, shape bias, robustness, and transferability to downstream tasks) with supervised teachers, self-supervised teaching assistants and students are inductive and may inspire future researches.

Channel pruning and tensor decomposition have received extensive attention in convolutional neural network compression. However, these two techniques are traditionally deployed in an isolated manner, leading to significant accuracy drop when pursuing high compression rates. In this paper, we propose a Collaborative Compression (CC) scheme, which joints channel pruning and tensor decomposition to compress CNN models by simultaneously learning the model sparsity and low-rankness. Specifically, we first investigate the compression sensitivity of each layer in the network, and then propose a Global Compression Rate Optimization that transforms the decision problem of compression rate into an optimization problem. After that, we propose multi-step heuristic compression to remove redundant compression units step-by-step, which fully considers the effect of the remaining compression space (i.e., unremoved compression units). Our method demonstrates superior performance gains over previous ones on various datasets and backbone architectures. For example, we achieve 52.9% FLOPs reduction by removing 48.4% parameters on ResNet-50 with only a Top-1 accuracy drop of 0.56% on ImageNet 2012.

The Information Bottleneck (IB) provides an information theoretic principle for representation learning, by retaining all information relevant for predicting label while minimizing the redundancy. Though IB principle has been applied to a wide range of applications, its optimization remains a challenging problem which heavily relies on the accurate estimation of mutual information. In this paper, we present a new strategy, Variational Self-Distillation (VSD), which provides a scalable, flexible and analytic solution to essentially fitting the mutual information but without explicitly estimating it. Under rigorously theoretical guarantee, VSD enables the IB to grasp the intrinsic correlation between representation and label for supervised training. Further-more, by extending VSD to multi-view learning, we introduce two other strategies, Variational Cross-Distillation (VCD) and Variational Mutual-Learning (VML), which significantly improve the robustness of representation to view-changes by eliminating view-specific and task-irrelevant in-formation. To verify our theoretically grounded strategies, we apply our approaches to cross-modal person Re-ID, and conduct extensive experiments, where the superior performance against state-of-the-art methods are demonstrated. Our intriguing findings highlight the need to rethink the way to estimate mutual information.

Single image dehazing is a challenging ill-posed problem due to the severe information degeneration. However, existing deep learning based dehazing methods only adopt clear images as positive samples to guide the training of dehazing network while negative information is unexploited. Moreover, most of them focus on strengthening the dehazing network with an increase of depth and width, leading to a significant requirement of computation and memory. In this paper, we propose a novel contrastive regularization (CR) built upon contrastive learning to exploit both the information of hazy images and clear images as negative and positive samples, respectively. CR ensures that the restored image is pulled to closer to the clear image and pushed to far away from the hazy image in the representation space.Furthermore, considering trade-off between performance and memory storage, we develop a compact dehazing network based on autoencoder-like (AE) framework. It involves an adaptive mixup operation and a dynamic feature enhancement module, which can benefit from preserving information flow adaptively and expanding the receptive field to improve the network’s transformation capability, respectively. We term our dehazing network with autoencoder and contrastive regularization as AECR-Net. The extensive experiments on synthetic and real-world datasets demonstrate that our AECR-Net surpass the state-of-the-art approaches. The code is released in https://github.com/GlassyWu/AECR-Net.

Novelty detection is the process of determining whether a query example differs from the learned training distribution. Previous generative adversarial networks based methods and self-supervised approaches suffer from instability training, mode dropping, and low discriminative ability. We overcome such problems by introducing a novel decoder-encoder framework. Firstly, a generative network (decoder) learns the representation by mapping the initialized latent vector to an image. In particular, this vector is initialized by considering the entire distribution of training data to avoid the problem of mode-dropping. Secondly, a contrastive network (encoder) aims to ``learn to compare'' through mutual information estimation, which directly helps the generative network to obtain a more discriminative representation by using a negative data augmentation strategy. Extensive experiments show that our model has significant superiority over cutting-edge novelty detectors and achieves new state-of-the-art results on various novelty detection benchmarks, e.g. CIFAR10 and DCASE. Moreover, our model is more stable for training in a non-adversarial manner, compared to other adversarial based novelty detection methods.

Human beings have a great generalization ability to recognize a novel category by only seeing a few number of samples. This is because humans possess the ability to learn from the concepts that already exist in our minds. However, many existing few-shot approaches fail in addressing such a fundamental problem, {\it i.e.,} how to utilize the knowledge learned in the past to improve the prediction for the new task. In this paper, we present a novel purified memory mechanism that simulates the recognition process of human beings. This new memory updating scheme enables the model to purify the information from semantic labels and progressively learn consistent, stable, and expressive concepts when episodes are trained one by one. On its basis, a Graph Augmentation Module (GAM) is introduced to aggregate these concepts and knowledge learned from new tasks via a graph neural network, making the prediction more accurate. Generally, our approach is model-agnostic and computing efficient with negligible memory cost. Extensive experiments performed on several benchmarks demonstrate the proposed method can consistently outperform a vast number of state-of-the-art few-shot learning methods.

Convolutional neural networks (CNNs) are highly successful for super-resolution (SR) but often require sophisticated architectures with heavy memory cost and computational overhead significantly restricts their practical deployments on resource-limited devices. In this paper, we proposed a novel contrastive self-distillation (CSD) framework to simultaneously compress and accelerate various off-the-shelf SR models. In particular, a channel-splitting super-resolution network can first be constructed from a target teacher network as a compact student network. Then, we propose a novel contrastive loss to improve the quality of SR images and PSNR/SSIM via explicit knowledge transfer. Extensive experiments demonstrate that the proposed CSD scheme effectively compresses and accelerates several standard SR models such as EDSR, RCAN and CARN. Code is available at https://github.com/Booooooooooo/CSD.

The success of convolutional neural networks (CNNs) in computer vision applications has been accompanied by a significant increase of computation and memory costs, which prohibits their usage on resource-limited environments, such as mobile systems or embedded devices. To this end, the research of CNN compression has recently become emerging. In this paper, we propose a novel filter pruning scheme, termed structured sparsity regularization (SSR), to simultaneously speed up the computation and reduce the memory overhead of CNNs, which can be well supported by various off-the-shelf deep learning libraries. Concretely, the proposed scheme incorporates two different regularizers of structured sparsity into the original objective function of filter pruning, which fully coordinates the global output and local pruning operations to adaptively prune filters. We further propose an alternative updating with Lagrange multipliers (AULM) scheme to efficiently solve its optimization. AULM follows the principle of alternating direction method of multipliers (ADMM) and alternates between promoting the structured sparsity of CNNs and optimizing the recognition loss, which leads to a very efficient solver ( 2.5× to the most recent work that directly solves the group sparsity-based regularization). Moreover, by imposing the structured sparsity, the online inference is extremely memory-light since the number of filters and the output feature maps are simultaneously reduced. The proposed scheme has been deployed to a variety of state-of-the-art CNN structures, including LeNet, AlexNet, VGGNet, ResNet, and GoogLeNet, over different data sets. Quantitative results demonstrate that the proposed scheme achieves superior performance over the state-of-the-art methods. We further demonstrate the proposed compression scheme for the task of transfer learning, including domain adaptation and object detection, which also show exciting performance gains over the state-of-the-art filter pruning methods.

Compressing convolutional neural networks (CNNs) has received ever-increasing research focus. However, most existing CNN compression methods do not interpret their inherent structures to distinguish the implicit redundancy. In this paper, we investigate the problem of CNN compression from a novel interpretable perspective. The relationship between the input feature maps and 2D kernels is revealed in a theoretical framework, based on which a kernel sparsity and entropy (KSE) indicator is proposed to quantitate the feature map importance in a feature-agnostic manner to guide model compression. Kernel clustering is further conducted based on the KSE indicator to accomplish high-precision CNN compression. KSE is capable of simultaneously compressing each layer in an efficient way, which is significantly faster compared to previous data-driven feature map pruning methods. We comprehensively evaluate the compression and speedup of the proposed method on CIFAR-10, SVHN and ImageNet 2012. Our method demonstrates superior performance gains over previous ones. In particular, it achieves 4.7× FLOPs reduction and 2.9× compression on ResNet-50 with only a top-5 accuracy drop of 0.35% on ImageNet 2012, which significantly outperforms state-of-the-art methods.

Structured pruning of filters or neurons has received increased focus for compressing convolutional neural networks. Most existing methods rely on multi-stage optimizations in a layer-wise manner for iteratively pruning and retraining which may not be optimal and may be computation intensive. Besides, these methods are designed for pruning a specific structure, such as filter or block structures without jointly pruning heterogeneous structures. In this paper, we propose an effective structured pruning approach that jointly prunes filters as well as other structures in an end-to-end manner. To accomplish this, we first introduce a soft mask to scale the output of these structures by defining a new objective function with sparsity regularization to align the output of baseline and network with this mask. We then effectively solve the optimization problem by generative adversarial learning (GAL), which learns a sparse soft mask in a label-free and an end-to-end manner. By forcing more scale factors in the soft mask to zero, the fast iterative shrinkage-thresholding algorithm (FISTA) can be leveraged to fast and reliably remove the corresponding structures. Extensive experiments demonstrate the effectiveness of GAL on different datasets, including MNIST, CIFAR-10 and ImageNet ILSVRC 2012. For example, on ImageNet ILSVRC 2012, the pruned ResNet-50 achieves 10.88% Top-5 error and results in a factor of 3.7x speedup. This significantly outperforms state-of-the-art methods.

Convolutional neural networks (CNNs) have achieved remarkable success in various computer vision tasks, which are extremely powerful to deal with massive training data by using tens of millions of parameters. However, CNNs often cost significant memory and computation consumption, which prohibits their usage in resource-limited environments such as mobile or embedded devices. To address the above issues, the existing approaches typically focus on either accelerating the convolutional layers or compressing the fully-connected layers separatedly, without pursuing a joint optimum. In this paper, we overcome such a limitation by introducing a holistic CNN compression framework, termed LRDKT, which works throughout both convolutional and fully-connected layers. First, a low-rank decomposition (LRD) scheme is proposed to remove redundancies across both convolutional kernels and fully-connected matrices, which has a novel closed-form solver to significantly improve the efficiency of the existing iterative optimization solvers. Second, a novel knowledge transfer (KT) based training scheme is introduced. To recover the accumulated accuracy loss and overcome the vanishing gradient, KT explicitly aligns outputs and intermediate responses from a teacher (original) network to its student (compressed) network. We have comprehensively analyzed and evaluated the compression and speedup ratios of the proposed model on MNIST and ILSVRC 2012 benchmarks. In both benchmarks, the proposed scheme has demonstrated superior performance gains over the state-of-the-art methods. We also demonstrate the proposed compression scheme for the task of transfer learning, including domain adaptation and object detection, which show exciting performance gains over the state-of-the-arts. Our source code and compressed models are available at https://github.com/ShaohuiLin/LRDKT.

Accelerating convolutional neural networks has recently received ever-increasing research focus. Among various approaches proposed in the literature, filter pruning has been regarded as a promising solution, which is due to its advantage in significant speedup and memory reduction of both network model and intermediate feature maps. To this end, most approaches tend to prune filters in a layer-wise fixed manner, which is incapable to dynamically recover the previously removed filter, as well as jointly optimize the pruned network across layers. In this paper, we propose a novel global & dynamic pruning (GDP) scheme to prune redundant filters for CNN acceleration. In particular, GDP first globally prunes the unsalient filters across all layers by proposing a global discriminative function based on prior knowledge of filters. Second, it dynamically updates the filter saliency all over the pruned sparse network, and then recover the mistakenly pruned filter, followed by a retraining phase to improve the model accuracy. Specially, we effectively solve the corresponding non-convex optimization problem of the proposed GDP via stochastic gradient descent with greedy alternative updating. Extensive experiments show that, comparing to the state-of-the-art filter pruning methods, the proposed approach achieves superior performance to accelerate several cutting-edge CNNs on the ILSVRC 2012 benchmark.

Recent years have witnessed an extensive popularity of convolutional neural networks (CNNs) in various computer vision and artificial intelligence applications. However, the performance gains have come at a cost of substantially intensive computation complexity, which prohibits its usage inresource-limited applications like mobile or embedded devices. While increasing attention has been paid to the acceleration of internal network structure, the redundancy of visual input is rarely considered. In this paper, we make the first attempt of reducing spatial and channel redundancy directly from the visual input for CNNs acceleration. The proposed method, termed ESPACE (Elimination of SPAtial and Channel rEdundancy), works by the following three steps: First, the 3D channel redundancy of convolutional layers is reduced by a set of low-rank approximation of convolutional filters. Second, a novel mask based selective processing scheme is proposed, which further speedups the convolution operations via skipping unsalient spatial locations of the visual input. Third, the accelerated network is fine-tuned using the training data via back-propagation. The proposed method is evaluated on ImageNet 2012 with implementations on two widely adopted CNNs, i.e. AlexNet and GoogLeNet. In comparison to several recent methods of CNN acceleration, the proposed scheme has demonstrated new state-of-the-art acceleration performance by a factor of 5.48* and 4.12* speedup on AlexNet and GoogLeNet, respectively, with a minimal decrease in classification accuracy.