Recognition Benchmark Research Articles

A large corpus for the recognition of Greek Sign Language gestures

Sign language recognition (SLR) from videos constitutes a captivating problem in gesture recognition, requiring the interpretation of hand movements, facial expressions, and body postures. The complexity of sign formation, signing variability among signers, and the technical hurdles of visual detection and tracking render SLR a challenging task. At the same time, the scarcity of large-scale SLR datasets, which are critical for developing robust data-intensive deep-learning SLR models, exacerbates these issues. In this article, we introduce a multi-signer video corpus of Greek Sign Language (GSL), which is the largest GSL database to date, serving as a valuable resource for SLR research. This corpus comprises an extensive RGB+D video collection that conveys rich lexical content in a multi-modal fashion, encompassing three subsets: (i) isolated signs; (ii) continuous signing; and (iii) continuous alphabet fingerspelling of words. Moreover, we introduce a comprehensive experimental setup that paves the way for more accurate and robust SLR solutions. In particular, except for the multi-signer (MS) and signer-independent (SI) settings, we employ a signer-adapted (SA) experimental paradigm, facilitating a comprehensive evaluation of system performance across various scenarios. Further, we provide three baseline SLR systems for isolated signs, continuous signing, and continuous fingerspelling. These systems leverage cutting-edge methods in deep learning and sequence modeling to capture the intricate temporal dynamics inherent in sign gestures. The models are evaluated on the three corpus subsets, setting their state-of-the-art recognition benchmark. The SL-ReDu GSL corpus, including its recommended experimental frameworks, is publicly available at https://sl-redu.e-ce.uth.gr/corpus.

XFMP: A Benchmark for Explainable Fine-Grained Abnormal Behavior Recognition on Medical Personal Protective Equipment

XFMP: A Benchmark for Explainable Fine-Grained Abnormal Behavior Recognition on Medical Personal Protective Equipment

Cross-view action recognition understanding from exocentric to egocentric perspective

Understanding action recognition in egocentric videos has emerged as a vital research topic with numerous practical applications. With the limitation in the scale of egocentric data collection, learning robust deep learning-based action recognition models remains difficult. Transferring knowledge learned from the large-scale exocentric data to the egocentric data is challenging due to the difference in videos across views. Our work introduces a novel cross-view learning approach to action recognition (CVAR) that effectively transfers knowledge from the exocentric to the selfish view. First, we present a novel geometric-based constraint into the self-attention mechanism in Transformer based on analyzing the camera positions between two views. Then, we propose a new cross-view self-attention loss learned on unpaired cross-view data to enforce the self-attention mechanism learning to transfer knowledge across views. Finally, to further improve the performance of our cross-view learning approach, we present the metrics to measure the correlations in videos and attention maps effectively. Experimental results on standard egocentric action recognition benchmarks, i.e., Charades-Ego, EPIC-Kitchens-55, and EPIC-Kitchens-100, have shown our approach’s effectiveness and state-of-the-art performance.

Dynamic Routing and Knowledge Re-Learning for Data-Free Black-Box Attack.

Deep learning models have emerged as strong and efficient tools that can be applied to a broad spectrum of complex learning problems and many real-world applications. However, more and more works show that deep models are vulnerable to adversarial examples. Compared to vanilla attack settings, this paper advocates a more practical setting of data-free black-box attack, for which the attackers can completely not access the structures and parameters of the target model, as well as the intermediate features and any training data associated with the model. To tackle this task, previous methods generate transferable adversarial examples from a transparent substitute model to the target model. However, we found that these works have the limitations of taking static substitute model structure for different targets, only using hard synthesized examples once, and still relying on data statistics of the target model. This may potentially harm the performance of attacking the target model. To this end, we propose a novel Dynamic Routing and Knowledge Re-Learning framework (DraKe) to effectively learn a dynamic substitute model from the target model. Specifically, given synthesized training samples, a dynamic substitute structure learning strategy is proposed to adaptively generate optimal substitute model structure via a policy network according to different target models and tasks. To facilitate the substitute training, we present a graph-based structure information learning to capture the structural knowledge learned from the target model. For the inherent limitation that online data generation can only be learned once, a dynamic knowledge re-learning strategy is proposed to adjust the weights of optimization objectives and re-learn hard samples. Extensive experiments on four public image classification datasets and one face recognition benchmark are conducted to evaluate the efficacy of our Drake. We can obtain significant improvement compared with state-of-the-art competitors. More importantly, our DraKe consistently achieves attack superiority for different target models (e.g., residual networks, and vision transformers), showing great potential for complex real-world applications.

InfoGCN++: Learning Representation by Predicting the Future for Online Skeleton-based Action Recognition.

Skeleton-based action recognition has made significant advancements recently, with models like InfoGCN showcasing remarkable accuracy. However, these models exhibit a key limitation: they necessitate complete action observation prior to classification, which constrains their applicability in realtime situations such as surveillance and robotic systems. To overcome this barrier, we introduce InfoGCN++, an innovative extension of InfoGCN, explicitly developed for online skeletonbased action recognition. InfoGCN++ augments the abilities of the original InfoGCN model by allowing real-time categorization of action types, independent of the observation sequence's length. It transcends conventional approaches by learning from current and anticipated future movements, thereby creating a more thorough representation of the entire sequence. Our approach to prediction is managed as an extrapolation issue, grounded on observed actions. To enable this, InfoGCN++ incorporates Neural Ordinary Differential Equations, a concept that lets it effectively model the continuous evolution of hidden states. Following rigorous evaluations on three skeleton-based action recognition benchmarks, InfoGCN++ demonstrates exceptional performance in online action recognition. It consistently equals or exceeds existing techniques, highlighting its significant potential to reshape the landscape of real-time action recognition applications. Consequently, this work represents a major leap forward from InfoGCN, pushing the limits of what's possible in online, skeleton-based action recognition.

HunFlair2 in a cross-corpus evaluation of biomedical named entity recognition and normalization tools.

With the exponential growth of the life sciences literature, biomedical text mining (BTM) has become an essential technology for accelerating the extraction of insights from publications. The identification of entities in texts, such as diseases or genes, and their normalization, i.e. grounding them in knowledge base, are crucial steps in any BTM pipeline to enable information aggregation from multiple documents. However, tools for these two steps are rarely applied in the same context in which they were developed. Instead, they are applied "in the wild," i.e. on application-dependent text collections from moderately to extremely different from those used for training, varying, e.g. in focus, genre or text type. This raises the question whether the reported performance, usually obtained by training and evaluating on different partitions of the same corpus, can be trusted for downstream applications. Here, we report on the results of a carefully designed cross-corpus benchmark for entity recognition and normalization, where tools were applied systematically to corpora not used during their training. Based on a survey of 28 published systems, we selected five, based on predefined criteria like feature richness and availability, for an in-depth analysis on three publicly available corpora covering four entity types. Our results present a mixed picture and show that cross-corpus performance is significantly lower than the in-corpus performance. HunFlair2, the redesigned and extended successor of the HunFlair tool, showed the best performance on average, being closely followed by PubTator Central. Our results indicate that users of BTM tools should expect a lower performance than the original published one when applying tools in "the wild" and show that further research is necessary for more robust BTM tools. All our models are integrated into the Natural Language Processing (NLP) framework flair: https://github.com/flairNLP/flair. Code to reproduce our results is available at: https://github.com/hu-ner/hunflair2-experiments.

MAFormer: A cross-channel spatio-temporal feature aggregation method for human action recognition

Human action recognition has been widely used in fields such as human–computer interaction and virtual reality. Despite significant progress, existing approaches still struggle with effectively integrating hierarchical information and processing data beyond a certain frame count. To address these challenges, we introduce the Multi-AxisFormer (MAFormer) model, which is organized in terms of spatial, temporal, and channel dimensions of the action sequence, thereby enhancing the model’s understanding of correlations and intricate structures among and within features. Drawing on the Transformer architecture, we propose the Cross-channel Spatio-temporal Aggregation (CSA) structure for more refined feature extraction and the Multi-Axis Attention (MAA) module for more comprehensive feature aggregation. Moreover, the integration of Rotary Position Embedding (RoPE) boosts the model’s extrapolation and generalization abilities. MAFormer surpasses the known state-of-the-art on multiple skeleton-based action recognition benchmarks with the accuracy of 93.2% on NTU RGB+D 60 cross-subject split, 89.9% on NTU RGB+D 120 cross-subject split, and 97.2% on N-UCLA, offering a novel paradigm for hierarchical modeling in human action recognition.

BViT: Broad Attention-Based Vision Transformer.

Recent works have demonstrated that transformer can achieve promising performance in computer vision, by exploiting the relationship among image patches with self-attention. They only consider the attention in a single feature layer, but ignore the complementarity of attention in different layers. In this article, we propose broad attention to improve the performance by incorporating the attention relationship of different layers for vision transformer (ViT), which is called BViT. The broad attention is implemented by broad connection and parameter-free attention. Broad connection of each transformer layer promotes the transmission and integration of information for BViT. Without introducing additional trainable parameters, parameter-free attention jointly focuses on the already available attention information in different layers for extracting useful information and building their relationship. Experiments on image classification tasks demonstrate that BViT delivers superior accuracy of 75.0%/81.6% top-1 accuracy on ImageNet with 5M/22M parameters. Moreover, we transfer BViT to downstream object recognition benchmarks to achieve 98.9% and 89.9% on CIFAR10 and CIFAR100, respectively, that exceed ViT with fewer parameters. For the generalization test, the broad attention in Swin Transformer, T2T-ViT and LVT also brings an improvement of more than 1%. To sum up, broad attention is promising to promote the performance of attention-based models. Code and pretrained models are available at https://github.com/DRL/BViT.

A hybrid transposed attention based deep learning model for wearable and explainable stress recognition

Stress is a prevalent issue in modern society which affects a large percentage of the population. Stress has negative effects on human daily lives such as reduced memory and concentration levels, forgetfulness, restlessness, and weakness. Although it is challenging to extinguish stress from human lives, monitoring and controlling its consequences can be implemented. Current methods for stress monitoring often lack interpretability of results and making it difficult to provide an interpretable system for immediate mental health counseling and intervention. Wearable devices are of growing interest due to the impact of the COVID-19 pandemic on mental health, as wearables provide monitoring of physiological parameters which helps in measuring stress levels. Here, an investigation on the feasibility of using wrist and chest-based multimodal physiological signals collected from wearables is executed for stress classification and identification of underlying modalities that impact the stress. Multimodalities from wrist and chest sensors are used to examine the three-class classification, baseline, stress and amusement. For improved and visualized effects, the multimodal signals from wearables were transformed into images. Then, a hybrid model was introduced, which comprised of attentive convolution neural network (CNN), transposed attentive CNN connections with long-short term memory (LSTM) and followed by attention layer for efficient feature extraction and improved classification accuracy. Results found that the proposed model with images provided 97 % accuracy and 96 % F1-score for chest wearable and 94 % accuracy and 93 % F1-score for wrist wearable data. Additionally, the post-hoc explainability approach (Local interpretable model agnostic applications, LIME) provided visual representations of contributing features from each signal for each class. LIME shows electrodermal activity (EDA), temperature (TEMP) and respiration (RESP) to be the significant factor in stress recognition from chest wearable and blood volume pulse (BVP) and EDA from wrist wearable. Additionally, increasing the number of features of the explainable model influences the modalities influence on the model explainability. The results establish a benchmark for explainable stress recognition using different sensor data.

EMS: A Large-Scale Eye Movement Dataset, Benchmark, and New Model for Schizophrenia Recognition.

Schizophrenia (SZ) is a common and disabling mental illness, and most patients encounter cognitive deficits. The eye-tracking technology has been increasingly used to characterize cognitive deficits for its reasonable time and economic costs. However, there is no large-scale and publicly available eye movement dataset and benchmark for SZ recognition. To address these issues, we release a large-scale Eye Movement dataset for SZ recognition (EMS), which consists of eye movement data from 104 schizophrenics and 104 healthy controls (HCs) based on the free-viewing paradigm with 100 stimuli. We also conduct the first comprehensive benchmark, which has been absent for a long time in this field, to compare the related 13 psychosis recognition methods using six metrics. Besides, we propose a novel mean-shift-based network (MSNet) for eye movement-based SZ recognition, which elaborately combines the mean shift algorithm with convolution to extract the cluster center as the subject feature. In MSNet, first, a stimulus feature branch (SFB) is adopted to enhance each stimulus feature with similar information from all stimulus features, and then, the cluster center branch (CCB) is utilized to generate the cluster center as subject feature and update it by the mean shift vector. The performance of our MSNet is superior to prior contenders, thus, it can act as a powerful baseline to advance subsequent study. To pave the road in this research field, the EMS dataset, the benchmark results, and the code of MSNet are publicly available at https://github.com/YingjieSong1/EMS.

SiSe: Simultaneous and Sequential Transformers for multi-label activity recognition

Multi-label activity recognition is extremely challenging, where multiple activities may appear simultaneously or sequentially in a video. While previous works have realized the temporal co-occurrence of activities, the sequential order of activities have been largely overlooked. However, we argue that the sequential order of activities should also be preserved in correlation modeling, because shuffling the order might not form a semantically meaningful video. In this work, we present plug-and-play Simultaneous and Sequential Transformer (SiSe) modules for multi-label activity recognition. Upon frame features of all time steps, SiSe enhances spatio-temporal feature learning for multi-label activity recognition, by capturing the simultaneous and sequential activity correlations. Specifically, we employ a Simultaneous Transformer module to connect multiple activities that probably appear at each frame, and a hierarchical Sequential Transformer module to efficiently capture the sequential activity correlations in an order-preserved manner. Despite the straightforward and class-agnostic design of SiSe, it can outperform state-of-the-art approaches on three multi-label activity recognition benchmarks. In particular, we verify the significance of preserving the sequential order of activities with our Sequential Transformer in correlation modeling. We also conduct ablation studies and visual analysis for better understanding of our SiSe.

Self-learning activation functions to increase accuracy of privacy-preserving Convolutional Neural Networks with homomorphic encryption.

The widespread adoption of cloud computing necessitates privacy-preserving techniques that allow information to be processed without disclosure. This paper proposes a method to increase the accuracy and performance of privacy-preserving Convolutional Neural Networks with Homomorphic Encryption (CNN-HE) by Self-Learning Activation Functions (SLAF). SLAFs are polynomials with trainable coefficients updated during training, together with synaptic weights, for each polynomial independently to learn task-specific and CNN-specific features. We theoretically prove its feasibility to approximate any continuous activation function to the desired error as a function of the SLAF degree. Two CNN-HE models are proposed: CNN-HE-SLAF and CNN-HE-SLAF-R. In the first model, all activation functions are replaced by SLAFs, and CNN is trained to find weights and coefficients. In the second one, CNN is trained with the original activation, then weights are fixed, activation is substituted by SLAF, and CNN is shortly re-trained to adapt SLAF coefficients. We show that such self-learning can achieve the same accuracy 99.38% as a non-polynomial ReLU over non-homomorphic CNNs and lead to an increase in accuracy (99.21%) and higher performance (6.26 times faster) than the state-of-the-art CNN-HE CryptoNets on the MNIST optical character recognition benchmark dataset.

Binary Encoding-Based Federated Learning for Traffic Sign Recognition in Autonomous Driving

Autonomous driving involves collaborative data sensing and traffic sign recognition. Emerging artificial intelligence technology has brought tremendous advances to vehicular networks. However, it is challenging to guarantee privacy and security when using traditional centralized machine learning methods for traffic sign recognition. It is urgent to introduce a distributed machine learning approach to protect private data of connected vehicles. In this paper, we propose a local differential privacy-based binary encoding federated learning approach. The binary encoding techniques and random perturbation methods are used in distributed learning scenarios to enhance the efficiency and security of data transmission. For the vehicle layer in this approach, the model is trained locally, and the model parameters are uploaded to the central server through encoding and perturbing. The central server designs the corresponding decoding, correction scheme, and regression statistical method for the received binary string. Then, the model parameters are aggregated and updated in the server and transmitted to the vehicle until the learning model is trained. The performance of the proposed approach is verified using the German Traffic Sign Recognition Benchmark data set. The simulation results show that the convergence of the approach is better with the increase in the learning cycle. Compared with baseline methods, such as the convolutional neural network, random forest, and backpropagation, the proposed approach achieves higher accuracy in the process of traffic sign recognition, with an increase of 6%.

TRAFFIC SIGN RECOGNITION IN CHALLENGING WEATHER CONDITIONS USING CONVOLUTIONAL NEURAL NETWORKS

The ability of autonomous driving systems to recognize traffic signs in a variety of environmental conditions is crucial to their reliability. This study uses convolutional neural networks (CNNs) to provide a novel method for improving the accuracy of traffic sign recognition systems under challenging weather situations. The research focuses on developing a CNN-based model that is trained using the augmented German Traffic Sign Recognition Benchmark (GTSRB) dataset, which contains over 50,000 labeled images of road signs across 43 categories. To attempt to overcome the limitations of traditional CNN models under adverse environmental conditions, this work presents adaptive feature extraction layers that are especially made to reduce visibility problems brought on by rain, fog, and snow. By taking a comprehensive approach, the model uses advanced data augmentation methods to simulate different weather scenarios, greatly increasing the diversity of the training dataset. Through an analysis of theoretical and practical aspects, the study demonstrates how CNNs enhance the accuracy and efficiency of road sign detection systems in a different weather condition. Research evaluates the model's effectiveness using metrics such as precision, recall, and F1-score, proving its capability to reduce false positives and accurately detect relevant road sign instances. Additionally, the paper highlights the importance of careful dataset preparation and augmentation, model optimization, and training enhancements to improve the performance of road sign detection systems. The results of this research offer benefits for intelligent transport systems, autonomous driving, and road safety, pointing towards further advancements in precise and dependable road sign recognition technology.

Perturbation Augmentation for Adversarial Training with Diverse Attacks

Adversarial Training (AT) aims to alleviate the vulnerability of deep neural networks to adversarial perturbations. However, the AT techniques struggle to maintain the performance on natural samples while improving the deep model’s robustness. The absence of perturbation diversity in generated during the adversarial training degrades the generalizability of the robust models, causing overfitting to particular perturbations and a decrease in natural performance. This study proposes an adversarial training framework that augments adversarial directions from a single-step attack to address the trade-off between robustness and generalization. Inspired by feature scattering adversarial training, the proposed framework computes a principal adversarial direction with a single-step attack that finds a perturbation disrupting the inter-sample relationships in the mini-batch during adversarial training. The principal direction obtained at each iteration is augmented by sampling new adversarial directions within a region spanning 45 degrees from the principal adversarial direction. The proposed adversarial training approach does not require extra backpropagation steps in adversarial direction augmentation. Therefore, generalization of the robust model is improved without posing an additional burden on the feature scattering adversarial training. Experiments on CIFAR-10, CIFAR-100, SVHN, Tiny-ImageNet, and The German Traffic Sign Recognition Benchmark consistently improve the accuracy on adversarial with an almost pristine natural performance.

Efficient pyramid channel attention network for pathological myopia recognition with pretraining-and-finetuning

Pathological myopia (PM) is the leading ocular disease for impaired vision worldwide. Clinically, the characteristics of pathology distribution in PM are global-local on the fundus image, which plays a significant role in assisting clinicians in diagnosing PM. However, most existing deep neural networks focused on designing complex architectures but rarely explored the pathology distribution prior of PM. To tackle this issue, we propose an efficient pyramid channel attention (EPCA) module, which fully leverages the potential of the clinical pathology prior of PM with pyramid pooling and multi-scale context fusion. Then, we construct EPCA-Net for automatic PM recognition based on fundus images by stacking a sequence of EPCA modules. Moreover, motivated by the recent pretraining-and-finetuning paradigm, we attempt to adapt pre-trained natural image models for PM recognition by freezing them and treating the EPCA and other attention modules as adapters. In addition, we construct a PM recognition benchmark termed PM-fundus by collecting fundus images of PM from publicly available datasets. The comprehensive experiments demonstrate the superiority of EPCA-Net over state-of-the-art methods in the PM recognition task. For example, EPCA-Net achieves 97.56% accuracy and outperforms ViT by 2.85% accuracy on the PM-fundus dataset. The results also show that our method based on the pretraining-and-finetuning paradigm achieves competitive performance through comparisons to part of previous methods based on traditional fine-tuning paradigm with fewer tunable parameters, which has the potential to leverage more natural image foundation models to address the PM recognition task in limited medical data regime.

An efficient motion visual learning method for video action recognition

Currently, efficient spatio-temporal information modeling is one of the key research components to solve the action recognition problem. Previous approaches focus on enhancing the backbone features individually using hierarchical structures, and unfortunately, most of them fail to achieve a better balance between the interactional adequacy of features within the structure. In this work, we propose an effective Multi-dimensional Adaptive Fusion Network (MDAF-Net), which can be embedded into the mainstream action recognition backbone in a plug-and-play manner to fully activate the transfer and representation of action features in the deep network. Specifically, our MDAF-Net contains two main components: the Adaptive Temporal Capture Module (ATCM) and the Extended Spatial and Channel Module (ESCM). The ATCM effectively suppresses the over-expression of similar features in adjacent frames and activates the expression of motion flow information. The ESCM further improves temporal modeling efficiency by extending the spatial feature perceptual field and enhancing channel attention. Extensive experiments on several challenging action recognition benchmarks, such as Something-Something V1&V2 and Kinetics-400, demonstrate that the proposed MDAF can achieve state-of-the-art and competitive performance.

Intrinsic structure exploitation with dual alignment for universal visual recognition

This research explores the complex yet applicable domain of universal visual recognition, where unlabeled datasets may encompass additional, as-of-yet unidentified classes and present a skewed feature distribution within labeled datasets. The primary challenge lies in concurrently distinguishing instances of known classes and uncovering new classes within the unlabeled data. This difficulty is compounded by two factors: category shift and distribution shift. Current methodologies typically overlook the potential of harnessing the inherent structural relationship between labeled and unlabeled datasets, and they fail to effectively isolate the semantic variability of the newly identified classes from the established ones. To overcome these hurdles, we introduce an innovative framework, which we have termed MESA (short for Manifold structure Exploitation with double Spaces dual Alignment), to advance the field of universal visual recognition. Our approach begins by identifying geometrically and semantically related pairs of nearest neighbors between labeled and unlabeled data, which serve as anchor points to foster an intrinsic match between the two datasets. By integrating an affinity score for similarity, we establish a novel cross-view, anchor-guided, self-supervised learning paradigm. This model is designed to spread the learned labels’ information and to conjoin novel semantic insights within the predictive space. To intensify the matching accuracy for the known classes and enhance the differentiation of novel classes, we further implement a confident-prototype-driven, self-supervised learning paradigm in a cross-view manner. This strategy is aimed at revealing the macroscopic category configuration within the embedding space. MESA employs a pioneering bidirectional dual-alignment mechanism that operates simultaneously in the embedding and prediction spaces, hence providing a more robust approach to confronting challenges brought about by distribution and category shifts. Our exhaustive evaluation across several extensive image recognition benchmarks substantiates MESA’s superior performance over the latest cutting-edge methods, marking a significant leap forward in the sphere of universal visual recognition. Finally, MESA is implemented in Python using the Pytorch machine-learning library, and it is freely available at https://github.com/aimeeyaoyao/MESA.

A lightweight network architecture for traffic sign recognition based on enhanced LeNet-5 network.

As an important part of the unmanned driving system, the detection and recognition of traffic sign need to have the characteristics of excellent recognition accuracy, fast execution speed and easy deployment. Researchers have applied the techniques of machine learning, deep learning and image processing to traffic sign recognition successfully. Considering the hardware conditions of the terminal equipment in the unmanned driving system, in this research work, the goal was to achieve a convolutional neural network (CNN) architecture that is lightweight and easily implemented for an embedded application and with excellent recognition accuracy and execution speed. As a classical CNN architecture, LeNet-5 network model was chosen to be improved, including image preprocessing, improving spatial pool convolutional neural network, optimizing neurons, optimizing activation function, etc. The test experiment of the improved network architecture was carried out on German Traffic Sign Recognition Benchmark (GTSRB) database. The experimental results show that the improved network architecture can obtain higher recognition accuracy in a short interference time, and the algorithm loss is significantly reduced with the progress of training. At the same time, compared with other lightweight network models, this network architecture gives a good recognition result, with a recognition accuracy of 97.53%. The network structure is simple, the algorithm complexity is low, and it is suitable for all kinds of terminal equipment, which can have a wider application in unmanned driving system.

Depression risk recognition based on gait: A benchmark

Recently, depression recognition has received considerable attention. Due to easy acquisition at a distance, gait-based depression recognition can be a useful tool for auxiliary diagnosis and self-help depression risk assessment. Most existing methods use a few hand-craft features for analysis. To further investigate the relationship between depression and gait, we collect a gait-based depression dataset named D-Gait. 27,120 gait sequences of 292 volunteers with informed consent are collected to support the data-driven methodology (the private information is eliminated). Multiple shooting angles and three kinds of clothing are taken into consideration. To the best of our knowledge, it is the first published gait-based depression dataset. Based on this dataset, a gait-based depression risk recognition benchmark is established. We systematically investigate representative methods with skeleton and silhouette data, thereby verifying most psychological research conclusions about the relationship between gait and depression. Besides, more instructive insights are given in our paper, which indicates the significant potential of gait-based depression risk recognition. The benchmark will advance the research on depression, and it can be accessed at https://github.com/BNU-IVC/D-Gait.