Search

Designing spectral convolutional networks is a formidable task in graph learning. In traditional spectral graph neural networks (GNNs), polynomial-based methods are commonly used to design filters via the Laplacian matrix. In practical applications, however, these polynomial methods encounter inherent limitations, which primarily arise from the the low-order truncation of polynomial filters and the lack of overall modeling of the graph spectrum. This leads to poor performance of existing spectral approaches on real-world graph data, especially when the spectrum is highly concentrated or contains many numerically identical values, as they tend to apply the exact same modulation to signals with the same frequencies. To overcome these issues, in this paper, we propose Graph State Space Network (GrassNet), a novel graph neural network with theoretical support that provides a simple yet effective scheme for designing and learning arbitrary graph spectral filters. In particular, our GrassNet introduces structured state space models (SSMs) to model the correlations of graph signals at different frequencies and derives a unique rectification for each frequency in the graph spectrum. To the best of our knowledge, our work is the first to employ SSMs for the design of GNN spectral filters, and it theoretically offers greater expressive power compared with polynomial filters. Extensive experiments on nine public benchmarks reveal that GrassNet achieves superior performance in real-world graph modeling tasks.

Graph neural networks are recognized for their strong performance across various applications, with the backpropagation algorithm playing a central role in the development of most GNN models. However, despite its effectiveness, BP has limitations that challenge its biological plausibility and affect the efficiency, scalability and parallelism of training neural networks for graph-based tasks. While several non-BP training algorithms, such as the direct feedback alignment, have been successfully applied to fully-connected and convolutional network components for handling Euclidean data, directly adapting these non-BP frameworks to manage non-Euclidean graph data in GNN models presents significant challenges. These challenges primarily arise from the violation of the i.i.d. assumption in graph data and the difficulty in accessing prediction errors for all samples (nodes) within the graph. To overcome these obstacles, in this paper we propose DFA-GNN, a novel forward learning framework tailored for GNNs with a case study of semi-supervised learning. The proposed method breaks the limitations of BP by using a dedicated forward training mechanism. Specifically, DFA-GNN extends the principles of DFA to adapt to graph data and unique architecture of GNNs, which incorporates the information of graph topology into the feedback links to accommodate the non-Euclidean characteristics of graph data. Additionally, for semi-supervised graph learning tasks, we developed a pseudo error generator that spreads residual errors from training data to create a pseudo error for each unlabeled node. These pseudo errors are then utilized to train GNNs using DFA. Extensive experiments on 10 public benchmarks reveal that our learning framework outperforms not only previous non-BP methods but also the standard BP methods, and it exhibits excellent robustness against various types of noise and attacks.

Although the performance of Temporal Action Segmentation (TAS) has improved in recent years, achieving promising results often comes with a high computational cost due to dense inputs, complex model structures, and resource-intensive post-processing requirements. To improve the efficiency while keeping the performance, we present a novel perspective centered on per-segment classification. By harnessing the capabilities of Transformers, we tokenize each video segment as an instance token, endowed with intrinsic instance segmentation. To realize efficient action segmentation, we introduce BaFormer, a boundary-aware Transformer network. It employs instance queries for instance segmentation and a global query for class-agnostic boundary prediction, yielding continuous segment proposals. During inference, BaFormer employs a simple yet effective voting strategy to classify boundary-wise segments based on instance segmentation. Remarkably, as a single-stage approach, BaFormer significantly reduces the computational costs, utilizing only 6% of the running time compared to state-of-the-art method DiffAct, while producing better or comparable accuracy over several popular benchmarks. The code for this project is publicly available at https://github.com/peiyao-w/BaFormer., Comment: 17 pages, 8 figures, 11 tables

Optical Doppler Tomography (ODT) is a blood flow imaging technique popularly used in bioengineering applications. The fundamental unit of ODT is the 1D frequency response along the A-line (depth), named raw A-scan. A 2D ODT image (B-scan) is obtained by first sensing raw A-scans along the B-line (width), and then constructing the B-scan from these raw A-scans via magnitude-phase analysis and post-processing. To obtain a high-resolution B-scan with a precise flow map, densely sampled A-scans are required in current methods, causing both computational and storage burdens. To address this issue, in this paper we propose a novel sparse reconstruction framework with four main sequential steps: 1) early magnitude-phase fusion that encourages rich interaction of the complementary information in magnitude and phase, 2) State Space Model (SSM)-based representation learning, inspired by recent successes in Mamba and VMamba, to naturally capture both the intra-A-scan sequential information and between-A-scan interactions, 3) an Inception-based Feedforward Network module (IncFFN) to further boost the SSM-module, and 4) a B-line Pixel Shuffle (BPS) layer to effectively reconstruct the final results. In the experiments on real-world animal data, our method shows clear effectiveness in reconstruction accuracy. As the first application of SSM for image reconstruction tasks, we expect our work to inspire related explorations in not only efficient ODT imaging techniques but also generic image enhancement., Comment: 19 pages, 5 figures

Textual backdoor attacks pose significant security threats. Current detection approaches, typically relying on intermediate feature representation or reconstructing potential triggers, are task-specific and less effective beyond sentence classification, struggling with tasks like question answering and named entity recognition. We introduce TABDet (Task-Agnostic Backdoor Detector), a pioneering task-agnostic method for backdoor detection. TABDet leverages final layer logits combined with an efficient pooling technique, enabling unified logit representation across three prominent NLP tasks. TABDet can jointly learn from diverse task-specific models, demonstrating superior detection efficacy over traditional task-specific methods., Comment: Findings of NAACL 2024

Localizing predefined 3D keypoints in a 2D image is an effective way to establish 3D-2D correspondences for 6DoF object pose estimation. However, unreliable localization results of invisible keypoints degrade the quality of correspondences. In this paper, we address this issue by localizing the important keypoints in terms of visibility. Since keypoint visibility information is currently missing in dataset collection process, we propose an efficient way to generate binary visibility labels from available object-level annotations, for keypoints of both asymmetric objects and symmetric objects. We further derive real-valued visibility-aware importance from binary labels based on PageRank algorithm. Taking advantage of the flexibility of our visibility-aware importance, we construct VAPO (Visibility-Aware POse estimator) by integrating the visibility-aware importance with a state-of-the-art pose estimation algorithm, along with additional positional encoding. Extensive experiments are conducted on popular pose estimation benchmarks including Linemod, Linemod-Occlusion, and YCB-V. The results show that, VAPO improves both the keypoint correspondences and final estimated poses, and clearly achieves state-of-the-art performances.

Motivated by the Parameter-Efficient Fine-Tuning (PEFT) in large language models, we propose LoRAT, a method that unveils the power of large ViT model for tracking within laboratory-level resources. The essence of our work lies in adapting LoRA, a technique that fine-tunes a small subset of model parameters without adding inference latency, to the domain of visual tracking. However, unique challenges and potential domain gaps make this transfer not as easy as the first intuition. Firstly, a transformer-based tracker constructs unshared position embedding for template and search image. This poses a challenge for the transfer of LoRA, usually requiring consistency in the design when applied to the pre-trained backbone, to downstream tasks. Secondly, the inductive bias inherent in convolutional heads diminishes the effectiveness of parameter-efficient fine-tuning in tracking models. To overcome these limitations, we first decouple the position embeddings in transformer-based trackers into shared spatial ones and independent type ones. The shared embeddings, which describe the absolute coordinates of multi-resolution images (namely, the template and search images), are inherited from the pre-trained backbones. In contrast, the independent embeddings indicate the sources of each token and are learned from scratch. Furthermore, we design an anchor-free head solely based on MLP to adapt PETR, enabling better performance with less computational overhead. With our design, 1) it becomes practical to train trackers with the ViT-g backbone on GPUs with only memory of 25.8GB (batch size of 16); 2) we reduce the training time of the L-224 variant from 35.0 to 10.8 GPU hours; 3) we improve the LaSOT SUC score from 0.703 to 0.742 with the L-224 variant; 4) we fast the inference speed of the L-224 variant from 52 to 119 FPS. Code and models are available at https://github.com/LitingLin/LoRAT., Comment: Accepted by ECCV 2024. All experiment results are updated

Single object tracking (SOT) is a fundamental problem in computer vision, with a wide range of applications, including autonomous driving, augmented reality, and robot navigation. The robustness of SOT faces two main challenges: tiny target and fast motion. These challenges are especially manifested in videos captured by unmanned aerial vehicles (UAV), where the target is usually far away from the camera and often with significant motion relative to the camera. To evaluate the robustness of SOT methods, we propose BioDrone -- the first bionic drone-based visual benchmark for SOT. Unlike existing UAV datasets, BioDrone features videos captured from a flapping-wing UAV system with a major camera shake due to its aerodynamics. BioDrone hence highlights the tracking of tiny targets with drastic changes between consecutive frames, providing a new robust vision benchmark for SOT. To date, BioDrone offers the largest UAV-based SOT benchmark with high-quality fine-grained manual annotations and automatically generates frame-level labels, designed for robust vision analyses. Leveraging our proposed BioDrone, we conduct a systematic evaluation of existing SOT methods, comparing the performance of 20 representative models and studying novel means of optimizing a SOTA method (KeepTrack KeepTrack) for robust SOT. Our evaluation leads to new baselines and insights for robust SOT. Moving forward, we hope that BioDrone will not only serve as a high-quality benchmark for robust SOT, but also invite future research into robust computer vision. The database, toolkits, evaluation server, and baseline results are available at http://biodrone.aitestunion.com., Comment: This paper is published in IJCV (refer to DOI). Please cite the published IJCV

Full projector compensation is a practical task of projector-camera systems. It aims to find a projector input image, named compensation image, such that when projected it cancels the geometric and photometric distortions due to the physical environment and hardware. State-of-the-art methods use deep learning to address this problem and show promising performance for low-resolution setups. However, directly applying deep learning to high-resolution setups is impractical due to the long training time and high memory cost. To address this issue, this paper proposes a practical full compensation solution. Firstly, we design an attention-based grid refinement network to improve geometric correction quality. Secondly, we integrate a novel sampling scheme into an end-to-end compensation network to alleviate computation and introduce attention blocks to preserve key features. Finally, we construct a benchmark dataset for high-resolution projector full compensation. In experiments, our method demonstrates clear advantages in both efficiency and quality.

It is important to quantify Damage Assessment (DA) for Human Assistance and Disaster Response (HADR) applications. In this paper, to achieve efficient and scalable DA in HADR, an image prior and posterior conditional probability (IP2CP) is developed as an effective computational imaging representation. Equipped with the IP2CP representation, the matching pre- and post-disaster images are effectively encoded into one image that is then processed using deep learning approaches to determine the damage levels. Two scenarios of crucial importance for the practical use of DA in HADR applications are examined: pixel-wise semantic segmentation and patch-based contrastive learning-based global damage classification. Results achieved by IP2CP in both scenarios demonstrate promising performances, showing that our IP2CP-based methods within the deep learning framework can effectively achieve data and computational efficiency, which is of utmost importance for the DA in HADR applications., Comment: 6 pages, 2 figures

Recent studies have revealed that \textit{Backdoor Attacks} can threaten the safety of natural language processing (NLP) models. Investigating the strategies of backdoor attacks will help to understand the model's vulnerability. Most existing textual backdoor attacks focus on generating stealthy triggers or modifying model weights. In this paper, we directly target the interior structure of neural networks and the backdoor mechanism. We propose a novel Trojan Attention Loss (TAL), which enhances the Trojan behavior by directly manipulating the attention patterns. Our loss can be applied to different attacking methods to boost their attack efficacy in terms of attack successful rates and poisoning rates. It applies to not only traditional dirty-label attacks, but also the more challenging clean-label attacks. We validate our method on different backbone models (BERT, RoBERTa, and DistilBERT) and various tasks (Sentiment Analysis, Toxic Detection, and Topic Classification)., Comment: Findings of EMNLP 2023

Existing methods for Salient Object Detection in Optical Remote Sensing Images (ORSI-SOD) mainly adopt Convolutional Neural Networks (CNNs) as the backbone, such as VGG and ResNet. Since CNNs can only extract features within certain receptive fields, most ORSI-SOD methods generally follow the local-to-contextual paradigm. In this paper, we propose a novel Global Extraction Local Exploration Network (GeleNet) for ORSI-SOD following the global-to-local paradigm. Specifically, GeleNet first adopts a transformer backbone to generate four-level feature embeddings with global long-range dependencies. Then, GeleNet employs a Direction-aware Shuffle Weighted Spatial Attention Module (D-SWSAM) and its simplified version (SWSAM) to enhance local interactions, and a Knowledge Transfer Module (KTM) to further enhance cross-level contextual interactions. D-SWSAM comprehensively perceives the orientation information in the lowest-level features through directional convolutions to adapt to various orientations of salient objects in ORSIs, and effectively enhances the details of salient objects with an improved attention mechanism. SWSAM discards the direction-aware part of D-SWSAM to focus on localizing salient objects in the highest-level features. KTM models the contextual correlation knowledge of two middle-level features of different scales based on the self-attention mechanism, and transfers the knowledge to the raw features to generate more discriminative features. Finally, a saliency predictor is used to generate the saliency map based on the outputs of the above three modules. Extensive experiments on three public datasets demonstrate that the proposed GeleNet outperforms relevant state-of-the-art methods. The code and results of our method are available at https://github.com/MathLee/GeleNet., Comment: 13 pages, 6 figures, Accepted by IEEE Transactions on Image Processing 2023

Cholecystectomy (gallbladder removal) is one of the most common procedures in the US, with more than 1.2M procedures annually. Compared with classical open cholecystectomy, laparoscopic cholecystectomy (LC) is associated with significantly shorter recovery period, and hence is the preferred method. However, LC is also associated with an increase in bile duct injuries (BDIs), resulting in significant morbidity and mortality. The primary cause of BDIs from LCs is misidentification of the cystic duct with the bile duct. Critical view of safety (CVS) is the most effective of safety protocols, which is said to be achieved during the surgery if certain criteria are met. However, due to suboptimal understanding and implementation of CVS, the BDI rates have remained stable over the last three decades. In this paper, we develop deep-learning techniques to automate the assessment of CVS in LCs. An innovative aspect of our research is on developing specialized learning techniques by incorporating domain knowledge to compensate for the limited training data available in practice. In particular, our CVS assessment process involves a fusion of two segmentation maps followed by an estimation of a certain region of interest based on anatomical structures close to the gallbladder, and then finally determination of each of the three CVS criteria via rule-based assessment of structural information. We achieved a gain of over 11.8% in mIoU on relevant classes with our two-stream semantic segmentation approach when compared to a single-model baseline, and 1.84% in mIoU with our proposed Sobel loss function when compared to a Transformer-based baseline model. For CVS criteria, we achieved up to 16% improvement and, for the overall CVS assessment, we achieved 5% improvement in balanced accuracy compared to DeepCVS under the same experiment settings.

Accurate tracking of transparent objects, such as glasses, plays a critical role in many robotic tasks such as robot-assisted living. Due to the adaptive and often reflective texture of such objects, traditional tracking algorithms that rely on general-purpose learned features suffer from reduced performance. Recent research has proposed to instill transparency awareness into existing general object trackers by fusing purpose-built features. However, with the existing fusion techniques, the addition of new features causes a change in the latent space making it impossible to incorporate transparency awareness on trackers with fixed latent spaces. For example, many of the current days transformer-based trackers are fully pre-trained and are sensitive to any latent space perturbations. In this paper, we present a new feature fusion technique that integrates transparency information into a fixed feature space, enabling its use in a broader range of trackers. Our proposed fusion module, composed of a transformer encoder and an MLP module, leverages key query-based transformations to embed the transparency information into the tracking pipeline. We also present a new two-step training strategy for our fusion module to effectively merge transparency features. We propose a new tracker architecture that uses our fusion techniques to achieve superior results for transparent object tracking. Our proposed method achieves competitive results with state-of-the-art trackers on TOTB, which is the largest transparent object tracking benchmark recently released. Our results and the implementation of code will be made publicly available at https://github.com/kalyan0510/TOTEM., Comment: IEEE IROS 2023

Domain Generalization (DG) aims to learn a generalizable model on the unseen target domain by only training on the multiple observed source domains. Although a variety of DG methods have focused on extracting domain-invariant features, the domain-specific class-relevant features have attracted attention and been argued to benefit generalization to the unseen target domain. To take into account the class-relevant domain-specific information, in this paper we propose an Information theory iNspired diSentanglement and pURification modEl (INSURE) to explicitly disentangle the latent features to obtain sufficient and compact (necessary) class-relevant feature for generalization to the unseen domain. Specifically, we first propose an information theory inspired loss function to ensure the disentangled class-relevant features contain sufficient class label information and the other disentangled auxiliary feature has sufficient domain information. We further propose a paired purification loss function to let the auxiliary feature discard all the class-relevant information and thus the class-relevant feature will contain sufficient and compact (necessary) class-relevant information. Moreover, instead of using multiple encoders, we propose to use a learnable binary mask as our disentangler to make the disentanglement more efficient and make the disentangled features complementary to each other. We conduct extensive experiments on four widely used DG benchmark datasets including PACS, OfficeHome, TerraIncognita, and DomainNet. The proposed INSURE outperforms the state-of-art methods. We also empirically show that domain-specific class-relevant features are beneficial for domain generalization., Comment: 10 pages, 4 figures

Multimodal vision-language (VL) learning has noticeably pushed the tendency toward generic intelligence owing to emerging large foundation models. However, tracking, as a fundamental vision problem, surprisingly enjoys less bonus from recent flourishing VL learning. We argue that the reasons are two-fold: the lack of large-scale vision-language annotated videos and ineffective vision-language interaction learning of current works. These nuisances motivate us to design more effective vision-language representation for tracking, meanwhile constructing a large database with language annotation for model learning. Particularly, in this paper, we first propose a general attribute annotation strategy to decorate videos in six popular tracking benchmarks, which contributes a large-scale vision-language tracking database with more than 23,000 videos. We then introduce a novel framework to improve tracking by learning a unified-adaptive VL representation, where the cores are the proposed asymmetric architecture search and modality mixer (ModaMixer). To further improve VL representation, we introduce a contrastive loss to align different modalities. To thoroughly evidence the effectiveness of our method, we integrate the proposed framework on three tracking methods with different designs, i.e., the CNN-based SiamCAR, the Transformer-based OSTrack, and the hybrid structure TransT. The experiments demonstrate that our framework can significantly improve all baselines on six benchmarks. Besides empirical results, we theoretically analyze our approach to show its rationality. By revealing the potential of VL representation, we expect the community to divert more attention to VL tracking and hope to open more possibilities for future tracking with diversified multimodal messages., Comment: 16 pages, 9 figures

Egocentric action recognition is gaining significant attention in the field of human action recognition. In this paper, we address data scarcity issue in egocentric action recognition from a compositional generalization perspective. To tackle this problem, we propose a free-form composition network (FFCN) that can simultaneously learn disentangled verb, preposition, and noun representations, and then use them to compose new samples in the feature space for rare classes of action videos. First, we use a graph to capture the spatial-temporal relations among different hand/object instances in each action video. We thus decompose each action into a set of verb and preposition spatial-temporal representations using the edge features in the graph. The temporal decomposition extracts verb and preposition representations from different video frames, while the spatial decomposition adaptively learns verb and preposition representations from action-related instances in each frame. With these spatial-temporal representations of verbs and prepositions, we can compose new samples for those rare classes in a free-form manner, which is not restricted to a rigid form of a verb and a noun. The proposed FFCN can directly generate new training data samples for rare classes, hence significantly improve action recognition performance. We evaluated our method on three popular egocentric action recognition datasets, Something-Something V2, H2O, and EPIC-KITCHENS-100, and the experimental results demonstrate the effectiveness of the proposed method for handling data scarcity problems, including long-tailed and few-shot egocentric action recognition.

Fully supervised action segmentation works on frame-wise action recognition with dense annotations and often suffers from the over-segmentation issue. Existing works have proposed a variety of solutions such as boundary-aware networks, multi-stage refinement, and temporal smoothness losses. However, most of them take advantage of frame-wise supervision, which cannot effectively tackle the evaluation metrics with different granularities. In this paper, for the desirable large receptive field, we first develop a novel local-global attention mechanism with temporal pyramid dilation and temporal pyramid pooling for efficient multi-scale attention. Then we decouple two inherent goals in action segmentation, ie, (1) individual identification solved by frame-wise supervision, and (2) temporal reasoning tackled by action set prediction. Afterward, an action alignment module fuses these different granularity predictions, leading to more accurate and smoother action segmentation. We achieve state-of-the-art accuracy, eg, 82.8% (+2.6%) on GTEA and 74.7% (+1.2%) on Breakfast, which demonstrates the effectiveness of our proposed method, accompanied by extensive ablation studies. The code will be made available later.

Estimating the 6-DoF pose of a rigid object from a single RGB image is a crucial yet challenging task. Recent studies have shown the great potential of dense correspondence-based solutions, yet improvements are still needed to reach practical deployment. In this paper, we propose a novel pose estimation algorithm named CheckerPose, which improves on three main aspects. Firstly, CheckerPose densely samples 3D keypoints from the surface of the 3D object and finds their 2D correspondences progressively in the 2D image. Compared to previous solutions that conduct dense sampling in the image space, our strategy enables the correspondence searching in a 2D grid (i.e., pixel coordinate). Secondly, for our 3D-to-2D correspondence, we design a compact binary code representation for 2D image locations. This representation not only allows for progressive correspondence refinement but also converts the correspondence regression to a more efficient classification problem. Thirdly, we adopt a graph neural network to explicitly model the interactions among the sampled 3D keypoints, further boosting the reliability and accuracy of the correspondences. Together, these novel components make CheckerPose a strong pose estimation algorithm. When evaluated on the popular Linemod, Linemod-O, and YCB-V object pose estimation benchmarks, CheckerPose clearly boosts the accuracy of correspondence-based methods and achieves state-of-the-art performances. Code is available at https://github.com/RuyiLian/CheckerPose., Comment: Accepted by ICCV2023

Deep neural networks are vulnerable to backdoor attacks, where an adversary maliciously manipulates the model behavior through overlaying images with special triggers. Existing backdoor defense methods often require accessing a few validation data and model parameters, which are impractical in many real-world applications, e.g., when the model is provided as a cloud service. In this paper, we address the practical task of blind backdoor defense at test time, in particular for black-box models. The true label of every test image needs to be recovered on the fly from a suspicious model regardless of image benignity. We focus on test-time image purification methods that incapacitate possible triggers while keeping semantic contents intact. Due to diverse trigger patterns and sizes, the heuristic trigger search in image space can be unscalable. We circumvent such barrier by leveraging the strong reconstruction power of generative models, and propose a framework of Blind Defense with Masked AutoEncoder (BDMAE). It detects possible triggers in the token space using image structural similarity and label consistency between the test image and MAE restorations. The detection results are then refined by considering trigger topology. Finally, we fuse MAE restorations adaptively into a purified image for making prediction. Our approach is blind to the model architectures, trigger patterns and image benignity. Extensive experiments under different backdoor settings validate its effectiveness and generalizability. Code is available at https://github.com/tsun/BDMAE.

Deep learning-based edge detectors heavily rely on pixel-wise labels which are often provided by multiple annotators. Existing methods fuse multiple annotations using a simple voting process, ignoring the inherent ambiguity of edges and labeling bias of annotators. In this paper, we propose a novel uncertainty-aware edge detector (UAED), which employs uncertainty to investigate the subjectivity and ambiguity of diverse annotations. Specifically, we first convert the deterministic label space into a learnable Gaussian distribution, whose variance measures the degree of ambiguity among different annotations. Then we regard the learned variance as the estimated uncertainty of the predicted edge maps, and pixels with higher uncertainty are likely to be hard samples for edge detection. Therefore we design an adaptive weighting loss to emphasize the learning from those pixels with high uncertainty, which helps the network to gradually concentrate on the important pixels. UAED can be combined with various encoder-decoder backbones, and the extensive experiments demonstrate that UAED achieves superior performance consistently across multiple edge detection benchmarks. The source code is available at \url{https://github.com/ZhouCX117/UAED}, Comment: CVPR2023

Gun violence is a critical security problem, and it is imperative for the computer vision community to develop effective gun detection algorithms for real-world scenarios, particularly in Closed Circuit Television (CCTV) surveillance data. Despite significant progress in visual object detection, detecting guns in real-world CCTV images remains a challenging and under-explored task. Firearms, especially handguns, are typically very small in size, non-salient in appearance, and often severely occluded or indistinguishable from other small objects. Additionally, the lack of principled benchmarks and difficulty collecting relevant datasets further hinder algorithmic development. In this paper, we present a meticulously crafted and annotated benchmark, called \textbf{CCTV-Gun}, which addresses the challenges of detecting handguns in real-world CCTV images. Our contribution is three-fold. Firstly, we carefully select and analyze real-world CCTV images from three datasets, manually annotate handguns and their holders, and assign each image with relevant challenge factors such as blur and occlusion. Secondly, we propose a new cross-dataset evaluation protocol in addition to the standard intra-dataset protocol, which is vital for gun detection in practical settings. Finally, we comprehensively evaluate both classical and state-of-the-art object detection algorithms, providing an in-depth analysis of their generalizing abilities. The benchmark will facilitate further research and development on this topic and ultimately enhance security. Code, annotations, and trained models are available at https://github.com/srikarym/CCTV-Gun.

Backpropagation algorithm has been widely used as a mainstream learning procedure for neural networks in the past decade, and has played a significant role in the development of deep learning. However, there exist some limitations associated with this algorithm, such as getting stuck in local minima and experiencing vanishing/exploding gradients, which have led to questions about its biological plausibility. To address these limitations, alternative algorithms to backpropagation have been preliminarily explored, with the Forward-Forward (FF) algorithm being one of the most well-known. In this paper we propose a new learning framework for neural networks, namely Cascaded Forward (CaFo) algorithm, which does not rely on BP optimization as that in FF. Unlike FF, our framework directly outputs label distributions at each cascaded block, which does not require generation of additional negative samples and thus leads to a more efficient process at both training and testing. Moreover, in our framework each block can be trained independently, so it can be easily deployed into parallel acceleration systems. The proposed method is evaluated on four public image classification benchmarks, and the experimental results illustrate significant improvement in prediction accuracy in comparison with the baseline.

Domain adaptive detection aims to improve the generalization of detectors on target domain. To reduce discrepancy in feature distributions between two domains, recent approaches achieve domain adaption through feature alignment in different granularities via adversarial learning. However, they neglect the relationship between multiple granularities and different features in alignment, degrading detection. Addressing this, we introduce a unified multi-granularity alignment (MGA)-based detection framework for domain-invariant feature learning. The key is to encode the dependencies across different granularities including pixel-, instance-, and category-levels simultaneously to align two domains. Specifically, based on pixel-level features, we first develop an omni-scale gated fusion (OSGF) module to aggregate discriminative representations of instances with scale-aware convolutions, leading to robust multi-scale detection. Besides, we introduce multi-granularity discriminators to identify where, either source or target domains, different granularities of samples come from. Note that, MGA not only leverages instance discriminability in different categories but also exploits category consistency between two domains for detection. Furthermore, we present an adaptive exponential moving average (AEMA) strategy that explores model assessments for model update to improve pseudo labels and alleviate local misalignment problem, boosting detection robustness. Extensive experiments on multiple domain adaption scenarios validate the superiority of MGA over other approaches on FCOS and Faster R-CNN detectors. Code will be released at https://github.com/tiankongzhang/MGA.

Due to the increasing computational demand of Deep Neural Networks (DNNs), companies and organizations have begun to outsource the training process. However, the externally trained DNNs can potentially be backdoor attacked. It is crucial to defend against such attacks, i.e., to postprocess a suspicious model so that its backdoor behavior is mitigated while its normal prediction power on clean inputs remain uncompromised. To remove the abnormal backdoor behavior, existing methods mostly rely on additional labeled clean samples. However, such requirement may be unrealistic as the training data are often unavailable to end users. In this paper, we investigate the possibility of circumventing such barrier. We propose a novel defense method that does not require training labels. Through a carefully designed layer-wise weight re-initialization and knowledge distillation, our method can effectively cleanse backdoor behaviors of a suspicious network with negligible compromise in its normal behavior. In experiments, we show that our method, trained without labels, is on-par with state-of-the-art defense methods trained using labels. We also observe promising defense results even on out-of-distribution data. This makes our method very practical. Code is available at: https://github.com/luluppang/BCU., Comment: Accepted to CVPR 2023

Active Domain Adaptation (ADA) queries the labels of a small number of selected target samples to help adapting a model from a source domain to a target domain. The local context of queried data is important, especially when the domain gap is large. However, this has not been fully explored by existing ADA works. In this paper, we propose a Local context-aware ADA framework, named LADA, to address this issue. To select informative target samples, we devise a novel criterion based on the local inconsistency of model predictions. Since the labeling budget is usually small, fine-tuning model on only queried data can be inefficient. We progressively augment labeled target data with the confident neighbors in a class-balanced manner. Experiments validate that the proposed criterion chooses more informative target samples than existing active selection strategies. Furthermore, our full method clearly surpasses recent ADA arts on various benchmarks. Code is available at https://github.com/tsun/LADA., Comment: Accepted by ICCV 2023

Enhancing model prediction confidence on target data is an important objective in Unsupervised Domain Adaptation (UDA). In this paper, we explore adversarial training on penultimate activations, i.e., input features of the final linear classification layer. We show that this strategy is more efficient and better correlated with the objective of boosting prediction confidence than adversarial training on input images or intermediate features, as used in previous works. Furthermore, with activation normalization commonly used in domain adaptation to reduce domain gap, we derive two variants and systematically analyze the effects of normalization on our adversarial training. This is illustrated both in theory and through empirical analysis on real adaptation tasks. Extensive experiments are conducted on popular UDA benchmarks under both standard setting and source-data free setting. The results validate that our method achieves the best scores against previous arts. Code is available at https://github.com/tsun/APA., Comment: AAAI 2023 Oral

Trojan attacks pose a severe threat to AI systems. Recent works on Transformer models received explosive popularity and the self-attentions are now indisputable. This raises a central question: Can we reveal the Trojans through attention mechanisms in BERTs and ViTs? In this paper, we investigate the attention hijacking pattern in Trojan AIs, \ie, the trigger token ``kidnaps'' the attention weights when a specific trigger is present. We observe the consistent attention hijacking pattern in Trojan Transformers from both Natural Language Processing (NLP) and Computer Vision (CV) domains. This intriguing property helps us to understand the Trojan mechanism in BERTs and ViTs. We also propose an Attention-Hijacking Trojan Detector (AHTD) to discriminate the Trojan AIs from the clean ones.

Contrastive language-image pretraining has shown great success in learning visual-textual joint representation from web-scale data, demonstrating remarkable "zero-shot" generalization ability for various image tasks. However, how to effectively expand such new language-image pretraining methods to video domains is still an open problem. In this work, we present a simple yet effective approach that adapts the pretrained language-image models to video recognition directly, instead of pretraining a new model from scratch. More concretely, to capture the long-range dependencies of frames along the temporal dimension, we propose a cross-frame attention mechanism that explicitly exchanges information across frames. Such module is lightweight and can be plugged into pretrained language-image models seamlessly. Moreover, we propose a video-specific prompting scheme, which leverages video content information for generating discriminative textual prompts. Extensive experiments demonstrate that our approach is effective and can be generalized to different video recognition scenarios. In particular, under fully-supervised settings, our approach achieves a top-1 accuracy of 87.1% on Kinectics-400, while using 12 times fewer FLOPs compared with Swin-L and ViViT-H. In zero-shot experiments, our approach surpasses the current state-of-the-art methods by +7.6% and +14.9% in terms of top-1 accuracy under two popular protocols. In few-shot scenarios, our approach outperforms previous best methods by +32.1% and +23.1% when the labeled data is extremely limited. Code and models are available at https://aka.ms/X-CLIP, Comment: Accepted by ECCV2022, Oral

Automatic action quality assessment (AQA) has attracted increasing attention due to its wide applications. However, most existing AQA methods employ deterministic models to predict the final score for each action, while overlooking the subjectivity and diversity among expert judges during the scoring process. In this paper, we propose a novel probabilistic model, named Uncertainty-Driven AQA (UD-AQA), to utilize and capture the diversity among multiple judge scores. Specifically, we design a Conditional Variational Auto-Encoder (CVAE)-based module to encode the uncertainty in expert assessment, where multiple judge scores can be produced by sampling latent features from the learned latent space multiple times. To further utilize the uncertainty, we generate the estimation of uncertainty for each prediction, which is employed to re-weight AQA regression loss, effectively reducing the influence of uncertain samples during training. Moreover, we further design an uncertainty-guided training strategy to dynamically adjust the learning order of the samples from low uncertainty to high uncertainty. The experiments show that our proposed method achieves competitive results on three benchmarks including the Olympic events MTL-AQA and FineDiving, and the surgical skill JIGSAWS datasets.

The waive of labels in the target domain makes Unsupervised Domain Adaptation (UDA) an attractive technique in many real-world applications, though it also brings great challenges as model adaptation becomes harder without labeled target data. In this paper, we address this issue by seeking compensation from target domain prior knowledge, which is often (partially) available in practice, e.g., from human expertise. This leads to a novel yet practical setting where in addition to the training data, some prior knowledge about the target class distribution are available. We term the setting as Knowledge-guided Unsupervised Domain Adaptation (KUDA). In particular, we consider two specific types of prior knowledge about the class distribution in the target domain: Unary Bound that describes the lower and upper bounds of individual class probabilities, and Binary Relationship that describes the relations between two class probabilities. We propose a general rectification module that uses such prior knowledge to refine model generated pseudo labels. The module is formulated as a Zero-One Programming problem derived from the prior knowledge and a smooth regularizer. It can be easily plugged into self-training based UDA methods, and we combine it with two state-of-the-art methods, SHOT and DINE. Empirical results on four benchmarks confirm that the rectification module clearly improves the quality of pseudo labels, which in turn benefits the self-training stage. With the guidance from prior knowledge, the performances of both methods are substantially boosted. We expect our work to inspire further investigations in integrating prior knowledge in UDA. Code is available at https://github.com/tsun/KUDA., Comment: To appear in ECCV 2022

Automatic recognition of surgical phases in surgical videos is a fundamental task in surgical workflow analysis. In this report, we propose a Transformer-based method that utilizes calibrated confidence scores for a 2-stage inference pipeline, which dynamically switches between a baseline model and a separately trained transition model depending on the calibrated confidence level. Our method outperforms the baseline model on the Cholec80 dataset, and can be applied to a variety of action segmentation methods.

In a data-driven paradigm, machine learning (ML) is the central component for developing accurate and universal exchange-correlation (XC) functionals in density functional theory (DFT). It is well known that XC functionals must satisfy several exact conditions and physical constraints, such as density scaling, spin scaling, and derivative discontinuity. In this work, we demonstrate that contrastive learning is a computationally efficient and flexible method to incorporate a physical constraint in ML-based density functional design. We propose a schematic approach to incorporate the uniform density scaling property of electron density for exchange energies by adopting contrastive representation learning during the pretraining task. The pretrained hidden representation is transferred to the downstream task to predict the exchange energies calculated by DFT. The electron density encoder transferred from the pretraining task based on contrastive learning predicts exchange energies that satisfy the scaling property, while the model trained without using contrastive learning gives poor predictions for the scaling-transformed electron density systems. Furthermore, the model with pretrained encoder gives a satisfactory performance with only small fractions of the whole augmented dataset labeled, comparable to the model trained from scratch using the whole dataset. The results demonstrate that incorporating exact constraints through contrastive learning can enhance the understanding of density-energy mapping using neural network (NN) models with less data labeling, which will be beneficial to generalizing the application of NN-based XC functionals in a wide range of scenarios that are not always available experimentally but theoretically justified. This work represents a viable pathway toward the machine learning design of a universal density functional via representation learning.

Unsupervised Domain Adaptation (UDA) aims to leverage a label-rich source domain to solve tasks on a related unlabeled target domain. It is a challenging problem especially when a large domain gap lies between the source and target domains. In this paper we propose a novel solution named SSRT (Safe Self-Refinement for Transformer-based domain adaptation), which brings improvement from two aspects. First, encouraged by the success of vision transformers in various vision tasks, we arm SSRT with a transformer backbone. We find that the combination of vision transformer with simple adversarial adaptation surpasses best reported Convolutional Neural Network (CNN)-based results on the challenging DomainNet benchmark, showing its strong transferable feature representation. Second, to reduce the risk of model collapse and improve the effectiveness of knowledge transfer between domains with large gaps, we propose a Safe Self-Refinement strategy. Specifically, SSRT utilizes predictions of perturbed target domain data to refine the model. Since the model capacity of vision transformer is large and predictions in such challenging tasks can be noisy, a safe training mechanism is designed to adaptively adjust learning configuration. Extensive evaluations are conducted on several widely tested UDA benchmarks and SSRT achieves consistently the best performances, including 85.43% on Office-Home, 88.76% on VisDA-2017 and 45.2% on DomainNet., Comment: To appear in CVPR 2022

The Chan-Vese (CV) model is a classic region-based method in image segmentation. However, its piecewise constant assumption does not always hold for practical applications. Many improvements have been proposed but the issue is still far from well solved. In this work, we propose an unsupervised image segmentation approach that integrates the CV model with deep neural networks, which significantly improves the original CV model's segmentation accuracy. Our basic idea is to apply a deep neural network that maps the image into a latent space to alleviate the violation of the piecewise constant assumption in image space. We formulate this idea under the classic Bayesian framework by approximating the likelihood with an evidence lower bound (ELBO) term while keeping the prior term in the CV model. Thus, our model only needs the input image itself and does not require pre-training from external datasets. Moreover, we extend the idea to multi-phase case and dataset based unsupervised image segmentation. Extensive experiments validate the effectiveness of our model and show that the proposed method is noticeably better than other unsupervised segmentation approaches.

Salient object detection (SOD) in optical remote sensing images (RSIs), or RSI-SOD, is an emerging topic in understanding optical RSIs. However, due to the difference between optical RSIs and natural scene images (NSIs), directly applying NSI-SOD methods to optical RSIs fails to achieve satisfactory results. In this paper, we propose a novel Adjacent Context Coordination Network (ACCoNet) to explore the coordination of adjacent features in an encoder-decoder architecture for RSI-SOD. Specifically, ACCoNet consists of three parts: an encoder, Adjacent Context Coordination Modules (ACCoMs), and a decoder. As the key component of ACCoNet, ACCoM activates the salient regions of output features of the encoder and transmits them to the decoder. ACCoM contains a local branch and two adjacent branches to coordinate the multi-level features simultaneously. The local branch highlights the salient regions in an adaptive way, while the adjacent branches introduce global information of adjacent levels to enhance salient regions. Additionally, to extend the capabilities of the classic decoder block (i.e., several cascaded convolutional layers), we extend it with two bifurcations and propose a Bifurcation-Aggregation Block to capture the contextual information in the decoder. Extensive experiments on two benchmark datasets demonstrate that the proposed ACCoNet outperforms 22 state-of-the-art methods under nine evaluation metrics, and runs up to 81 fps on a single NVIDIA Titan X GPU. The code and results of our method are available at https://github.com/MathLee/ACCoNet., Comment: 13 pages, 7 figures, Accepted by IEEE Transactions on Cybernetics 2022

Convolutional neural networks have made significant progresses in edge detection by progressively exploring the context and semantic features. However, local details are gradually suppressed with the enlarging of receptive fields. Recently, vision transformer has shown excellent capability in capturing long-range dependencies. Inspired by this, we propose a novel transformer-based edge detector, \emph{Edge Detection TransformER (EDTER)}, to extract clear and crisp object boundaries and meaningful edges by exploiting the full image context information and detailed local cues simultaneously. EDTER works in two stages. In Stage I, a global transformer encoder is used to capture long-range global context on coarse-grained image patches. Then in Stage II, a local transformer encoder works on fine-grained patches to excavate the short-range local cues. Each transformer encoder is followed by an elaborately designed Bi-directional Multi-Level Aggregation decoder to achieve high-resolution features. Finally, the global context and local cues are combined by a Feature Fusion Module and fed into a decision head for edge prediction. Extensive experiments on BSDS500, NYUDv2, and Multicue demonstrate the superiority of EDTER in comparison with state-of-the-arts., Comment: Accepted by CVPR2022

Optical coherence tomography angiography (OCTA) is an important imaging modality in many bioengineering tasks. The image quality of OCTA, however, is often degraded by Bulk Motion Artifacts (BMA), which are due to micromotion of subjects and typically appear as bright stripes surrounded by blurred areas. State-of-the-art methods usually treat BMA removal as a learning-based image inpainting problem, but require numerous training samples with nontrivial annotation. In addition, these methods discard the rich structural and appearance information carried in the BMA stripe region. To address these issues, in this paper we propose a self-supervised content-aware BMA removal model. First, the gradient-based structural information and appearance feature are extracted from the BMA area and injected into the model to capture more connectivity. Second, with easily collected defective masks, the model is trained in a self-supervised manner, in which only the clear areas are used for training while the BMA areas for inference. With the structural information and appearance feature from noisy image as references, our model can remove larger BMA and produce better visualizing result. In addition, only 2D images with defective masks are involved, hence improving the efficiency of our method. Experiments on OCTA of mouse cortex demonstrate that our model can remove most BMA with extremely large sizes and inconsistent intensities while previous methods fail.

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

Most previous learning-based graph matching algorithms solve the \textit{quadratic assignment problem} (QAP) by dropping one or more of the matching constraints and adopting a relaxed assignment solver to obtain sub-optimal correspondences. Such relaxation may actually weaken the original graph matching problem, and in turn hurt the matching performance. In this paper we propose a deep learning-based graph matching framework that works for the original QAP without compromising on the matching constraints. In particular, we design an affinity-assignment prediction network to jointly learn the pairwise affinity and estimate the node assignments, and we then develop a differentiable solver inspired by the probabilistic perspective of the pairwise affinities. Aiming to obtain better matching results, the probabilistic solver refines the estimated assignments in an iterative manner to impose both discrete and one-to-one matching constraints. The proposed method is evaluated on three popularly tested benchmarks (Pascal VOC, Willow Object and SPair-71k), and it outperforms all previous state-of-the-arts on all benchmarks.

Recently Transformer has been largely explored in tracking and shown state-of-the-art (SOTA) performance. However, existing efforts mainly focus on fusing and enhancing features generated by convolutional neural networks (CNNs). The potential of Transformer in representation learning remains under-explored. In this paper, we aim to further unleash the power of Transformer by proposing a simple yet efficient fully-attentional tracker, dubbed SwinTrack, within classic Siamese framework. In particular, both representation learning and feature fusion in SwinTrack leverage the Transformer architecture, enabling better feature interactions for tracking than pure CNN or hybrid CNN-Transformer frameworks. Besides, to further enhance robustness, we present a novel motion token that embeds historical target trajectory to improve tracking by providing temporal context. Our motion token is lightweight with negligible computation but brings clear gains. In our thorough experiments, SwinTrack exceeds existing approaches on multiple benchmarks. Particularly, on the challenging LaSOT, SwinTrack sets a new record with 0.713 SUC score. It also achieves SOTA results on other benchmarks. We expect SwinTrack to serve as a solid baseline for Transformer tracking and facilitate future research. Our codes and results are released at https://github.com/LitingLin/SwinTrack., Comment: 22 pages, 10 figures

In the computer vision community, great progresses have been achieved in salient object detection from natural scene images (NSI-SOD); by contrast, salient object detection in optical remote sensing images (RSI-SOD) remains to be a challenging emerging topic. The unique characteristics of optical RSIs, such as scales, illuminations and imaging orientations, bring significant differences between NSI-SOD and RSI-SOD. In this paper, we propose a novel Multi-Content Complementation Network (MCCNet) to explore the complementarity of multiple content for RSI-SOD. Specifically, MCCNet is based on the general encoder-decoder architecture, and contains a novel key component named Multi-Content Complementation Module (MCCM), which bridges the encoder and the decoder. In MCCM, we consider multiple types of features that are critical to RSI-SOD, including foreground features, edge features, background features, and global image-level features, and exploit the content complementarity between them to highlight salient regions over various scales in RSI features through the attention mechanism. Besides, we comprehensively introduce pixel-level, map-level and metric-aware losses in the training phase. Extensive experiments on two popular datasets demonstrate that the proposed MCCNet outperforms 23 state-of-the-art methods, including both NSI-SOD and RSI-SOD methods. The code and results of our method are available at https://github.com/MathLee/MCCNet., Comment: 12 pages, 7 figures, Accepted by IEEE Transactions on Geoscience and Remote Sensing 2021

Vision Transformer has shown great visual representation power in substantial vision tasks such as recognition and detection, and thus been attracting fast-growing efforts on manually designing more effective architectures. In this paper, we propose to use neural architecture search to automate this process, by searching not only the architecture but also the search space. The central idea is to gradually evolve different search dimensions guided by their E-T Error computed using a weight-sharing supernet. Moreover, we provide design guidelines of general vision transformers with extensive analysis according to the space searching process, which could promote the understanding of vision transformer. Remarkably, the searched models, named S3 (short for Searching the Search Space), from the searched space achieve superior performance to recently proposed models, such as Swin, DeiT and ViT, when evaluated on ImageNet. The effectiveness of S3 is also illustrated on object detection, semantic segmentation and visual question answering, demonstrating its generality to downstream vision and vision-language tasks. Code and models will be available at https://github.com/microsoft/Cream., Comment: Accepted to NIPS 2021

Objectives. The aim of this study was to investigate whether a deep convolutional neural network (CNN) with an attention module can detect osteoporosis on panoramic radiographs. Study Design. A dataset of 70 panoramic radiographs (PRs) from 70 different subjects of age between 49 to 60 was used, including 49 subjects with osteoporosis and 21 normal subjects. We utilized the leave-one-out cross-validation approach to generate 70 training and test splits. Specifically, for each split, one image was used for testing and the remaining 69 images were used for training. A deep convolutional neural network (CNN) using the Siamese architecture was implemented through a fine-tuning process to classify an PR image using patches extracted from eight representative trabecula bone areas (Figure 1). In order to automatically learn the importance of different PR patches, an attention module was integrated into the deep CNN. Three metrics, including osteoporosis accuracy (OPA), non-osteoporosis accuracy (NOPA) and overall accuracy (OA), were utilized for performance evaluation. Results. The proposed baseline CNN approach achieved the OPA, NOPA and OA scores of 0.667, 0.878 and 0.814, respectively. With the help of the attention module, the OPA, NOPA and OA scores were further improved to 0.714, 0.939 and 0.871, respectively. Conclusions. The proposed method obtained promising results using deep CNN with an attention module, which might be applied to osteoporosis prescreening., Comment: 9 pages

Many critical applications rely on cameras to capture video footage for analytical purposes. This has led to concerns about these cameras accidentally capturing more information than is necessary. In this paper, we propose a deep learning approach towards protecting privacy in camera-based systems. Instead of specifying specific objects (e.g. faces) are privacy sensitive, our technique distinguishes between salient (visually prominent) and non-salient objects based on the intuition that the latter is unlikely to be needed by the application.

In recent years, powered by the learned discriminative representation via graph neural network (GNN) models, deep graph matching methods have made great progresses in the task of matching semantic features. However, these methods usually rely on heuristically generated graph patterns, which may introduce unreliable relationships to hurt the matching performance. In this paper, we propose a joint \emph{graph learning and matching} network, named GLAM, to explore reliable graph structures for boosting graph matching. GLAM adopts a pure attention-based framework for both graph learning and graph matching. Specifically, it employs two types of attention mechanisms, self-attention and cross-attention for the task. The self-attention discovers the relationships between features and to further update feature representations over the learnt structures; and the cross-attention computes cross-graph correlations between the two feature sets to be matched for feature reconstruction. Moreover, the final matching solution is directly derived from the output of the cross-attention layer, without employing a specific matching decision module. The proposed method is evaluated on three popular visual matching benchmarks (Pascal VOC, Willow Object and SPair-71k), and it outperforms previous state-of-the-art graph matching methods by significant margins on all benchmarks. Furthermore, the graph patterns learnt by our model are validated to be able to remarkably enhance previous deep graph matching methods by replacing their handcrafted graph structures with the learnt ones.

While deep neural networks have shown impressive performance in many tasks, they are fragile to carefully designed adversarial attacks. We propose a novel adversarial training-based model by Attention Guided Knowledge Distillation and Bi-directional Metric Learning (AGKD-BML). The attention knowledge is obtained from a weight-fixed model trained on a clean dataset, referred to as a teacher model, and transferred to a model that is under training on adversarial examples (AEs), referred to as a student model. In this way, the student model is able to focus on the correct region, as well as correcting the intermediate features corrupted by AEs to eventually improve the model accuracy. Moreover, to efficiently regularize the representation in feature space, we propose a bidirectional metric learning. Specifically, given a clean image, it is first attacked to its most confusing class to get the forward AE. A clean image in the most confusing class is then randomly picked and attacked back to the original class to get the backward AE. A triplet loss is then used to shorten the representation distance between original image and its AE, while enlarge that between the forward and backward AEs. We conduct extensive adversarial robustness experiments on two widely used datasets with different attacks. Our proposed AGKD-BML model consistently outperforms the state-of-the-art approaches. The code of AGKD-BML will be available at: https://github.com/hongw579/AGKD-BML., Comment: ICCV 2021 paper