Video-based Action Recognition Research Articles

CANet: Comprehensive Attention Network for video-based action recognition

Attention mechanisms play a crucial role in improving action recognition performance. A video, a type of 3D data, can be effectively explored using attention mechanisms from temporal, spatial, and channel dimensions. However, existing methods based on 2D CNN tend to deal with complex spatiotemporal information from one or two of the dimensions, which eventually hampers their overall performance. In this paper, we propose a novel Comprehensive Attention Network (CANet) to model spatiotemporal information in all three dimensions adaptively. CANet is composed of three core plug-and-play components, namely the Global Guided Short-term Motion Module (GG-SMM), the Second-order Guided Long-term Motion Module (SG-LMM), and the Spatial Motion Adaptive Module (SMAM). Specifically, (1) the GG-SMM module is designed to represent local motion clues in the short-term temporal dimension to improve the classification accuracy of fast-tempo actions. (2) The SG-LMM module is designed to jointly motivate fine-grained motion information in the long-term temporal and channel dimensions, thereby facilitating the discrimination of long-term motions. (3) The SMAM module is used to represent motion-sensitive regions in the spatial dimension by learning the spatial object offsets. Extensive experiments have been conducted on four widely used action recognition benchmarks, namely, Something-Something V1, Kinetics-400, UCF-101, and HMDB-51. Experimental results demonstrate that the proposed CANet achieves excellent performance compared with other state-of-the-art methods.

Temporal-Distributed Backdoor Attack against Video Based Action Recognition

Deep neural networks (DNNs) have achieved tremendous success in various applications including video action recognition, yet remain vulnerable to backdoor attacks (Trojans). The backdoor-compromised model will mis-classify to the target class chosen by the attacker when a test instance (from a non-target class) is embedded with a specific trigger, while maintaining high accuracy on attack-free instances. Although there are extensive studies on backdoor attacks against image data, the susceptibility of video-based systems under backdoor attacks remains largely unexplored. Current studies are direct extensions of approaches proposed for image data, e.g., the triggers are independently embedded within the frames, which tend to be detectable by existing defenses. In this paper, we introduce a simple yet effective backdoor attack against video data. Our proposed attack, adding perturbations in a transformed domain, plants an imperceptible, temporally distributed trigger across the video frames, and is shown to be resilient to existing defensive strategies. The effectiveness of the proposed attack is demonstrated by extensive experiments with various well-known models on two video recognition benchmarks, UCF101 and HMDB51, and a sign language recognition benchmark, Greek Sign Language (GSL) dataset. We delve into the impact of several influential factors on our proposed attack and identify an intriguing effect termed "collateral damage" through extensive studies.

MgMViT: Multi-Granularity and Multi-Scale Vision Transformer for Efficient Action Recognition

Nowadays, the field of video-based action recognition is rapidly developing. Although Vision Transformers (ViT) have made great progress in static image processing, they are not yet fully optimized for dynamic video applications. Convolutional Neural Networks (CNN) and related models perform exceptionally well in video action recognition. However, there are still some issues that cannot be ignored, such as high computational costs and large memory consumption. In the face of these issues, current research focuses on finding effective methods to improve model performance and overcome current limits. Therefore, we present a unique Vision Transformer model based on multi-granularity and multi-scale fusion to accomplish efficient action recognition, which is designed for action recognition in videos to effectively reduce computational costs and memory usage. Firstly, we devise a multi-scale, multi-granularity module that integrates with Transformer blocks. Secondly, a hierarchical structure is utilized to manage information at various scales, and we introduce multi-granularity on top of multi-scale, which allows for a selective choice of the number of tokens to enter the next computational step, thereby reducing redundant tokens. Thirdly, a coarse-fine granularity fusion layer is introduced to reduce the sequence length of tokens with lower information content. The above two mechanisms are combined to optimize the allocation of resources in the model, further emphasizing critical information and reducing redundancy, thereby minimizing computational costs. To assess our proposed approach, comprehensive experiments are conducted by using benchmark datasets in the action recognition domain. The experimental results demonstrate that our method has achieved state-of-the-art performance in terms of accuracy and efficiency.

Unsupervised video-based action recognition using two-stream generative adversarial network

Unsupervised video-based action recognition using two-stream generative adversarial network

Knowledge-guided pre-training and fine-tuning: Video representation learning for action recognition

Video-based action recognition is an important task in the computer vision community, aiming to extract rich spatial–temporal information to recognize human actions from videos. Many approaches adopt self-supervised learning in large-scale unlabeled datasets and exploit transfer learning in the downstream action recognition task. Though much progress has been made for action recognition with video representation learning, two main issues remain for most existing methods. Firstly, the pre-training with self-supervised pretext tasks usually learns neutral and not much informative representations for the downstream action recognition task. Secondly, the valuable learned knowledge from large-scaled pre-training datasets will be gradually forgotten in the fine-tuning stage. To address such issues, in this paper, we propose a novel video representation learning method with knowledge-guided pre-training and fine-tuning for action recognition, which incorporates external human parsing knowledge for generating informative representation in the pre-training, and preserves the pre-trained knowledge in the fine-tuning stage to avoid catastrophic forgetting via self-distillation. Our model, with contributions from the external human parsing knowledge, video-level contrastive learning, and knowledge preserving self-distillation, achieves state-of-the-art performance on two popular benchmarks, i.e., UCF101 and HMDB51, verifying the effectiveness of the proposed method.

Action Recognition and Benchmark Using Event Cameras.

Recent years have witnessed remarkable achievements in video-based action recognition. Apart from traditional frame-based cameras, event cameras are bio-inspired vision sensors that only record pixel-wise brightness changes rather than the brightness value. However, little effort has been made in event-based action recognition, and large-scale public datasets are also nearly unavailable. In this paper,we present an event-based action recognition framework called EV-ACT. The Learnable Multi-Fused Representation (LMFR) is first proposed to integrate multiple event information in a learnable manner. The LMFR with dual temporal granularity is fed into the event-based slow-fast network for the fusion of appearance and motion features. A spatial-temporal attention mechanism is introduced to further enhance the learning capability of action recognition. To prompt research in this direction, we have collected the largest event-based action recognition benchmark named THUE-ACT-50 and the accompanying THUE-ACT-50-CHL dataset under challenging environments, including a total of over 12,830 recordings from 50 action categories, which is over 4 times the size of the previous largest dataset. Experimental results show that our proposed framework could achieve improvements of over 14.5%, 7.6%, 11.2%, and 7.4% compared to previous works on four benchmarks. We have also deployed our proposed EV-ACT framework on a mobile platform to validate its practicality and efficiency.

Towards efficient video-based action recognition: context-aware memory attention network

Abstract Given the prevalence of surveillance cameras in our daily lives, human action recognition from videos holds significant practical applications. A persistent challenge in this field is to develop more efficient models capable of real-time recognition with high accuracy for widespread implementation. In this research paper, we introduce a novel human action recognition model named Context-Aware Memory Attention Network (CAMA-Net), which eliminates the need for optical flow extraction and 3D convolution which are computationally intensive. By removing these components, CAMA-Net achieves superior efficiency compared to many existing approaches in terms of computation efficiency. A pivotal component of CAMA-Net is the Context-Aware Memory Attention Module, an attention module that computes the relevance score between key-value pairs obtained from the 2D ResNet backbone. This process establishes correspondences between video frames. To validate our method, we conduct experiments on four well-known action recognition datasets: ActivityNet, Diving48, HMDB51 and UCF101. The experimental results convincingly demonstrate the effectiveness of our proposed model, surpassing the performance of existing 2D-CNN based baseline models. Article Highlights Recent human action recognition models are not yet ready for practical applications due to high computation needs. We propose a 2D CNN-based human action recognition method to reduce the computation load. The proposed method achieves competitive performance compared to most SOTA 2D CNN-based methods on public datasets.

Temporal shuffling for defending deep action recognition models against adversarial attacks

Recently, video-based action recognition methods using convolutional neural networks (CNNs) achieve remarkable recognition performance. However, there is still lack of understanding about the generalization mechanism of action recognition models. In this paper, we suggest that action recognition models rely on the motion information less than expected, and thus they are robust to randomization of frame orders. Furthermore, we find that motion monotonicity remaining after randomization also contributes to such robustness. Based on this observation, we develop a novel defense method using temporal shuffling of input videos against adversarial attacks for action recognition models. Another observation enabling our defense method is that adversarial perturbations on videos are sensitive to temporal destruction. To the best of our knowledge, this is the first attempt to design a defense method without additional training for 3D CNN-based video action recognition models.

Video-Based Activity Recognition for Automated Motor Assessment of Parkinson's Disease.

Over the last decade, video-enabled mobile devices have become ubiquitous, while advances in markerless pose estimation allow an individual's body position to be tracked accurately and efficiently across the frames of a video. Previous work by this and other groups has shown that pose-extracted kinematic features can be used to reliably measure motor impairment in Parkinson's disease (PD). This presents the prospect of developing an asynchronous and scalable, video-based assessment of motor dysfunction. Crucial to this endeavour is the ability to automatically recognise the class of an action being performed, without which manual labelling is required. Representing the evolution of body joint locations as a spatio-temporal graph, we implement a deep-learning model for video and frame-level classification of activities performed according to part 3 of the Movement Disorder Society Unified PD Rating Scale (MDS-UPDRS). We train and validate this system using a dataset of n = 7310 video clips, recorded at 5 independent sites. This approach reaches human-level performance in detecting and classifying periods of activity within monocular video clips. Our framework could support clinical workflows and patient care at scale through applications such as quality monitoring of clinical data collection, automated labelling of video streams, or a module within a remote self-assessment system.

Efficient spatiotemporal context modeling for action recognition

Contextual information is essential in action recognition. However, local operations have difficulty in modeling two distant elements, and directly computing the dense relations between any two points brings huge computation and memory burden. Inspired by the recurrent 2D criss-cross attention (RCCA-2D) in image segmentation, we propose a recurrent 3D criss-cross attention (RCCA-3D) that factorizes the global relation map into sparse relation maps to model long-range spatiotemporal context with minor costs for video-based action recognition. Specifically, we first propose a 3D criss-cross attention (CCA-3D) module. Compared with the CCA-2D which only works in space, it can capture the spatiotemporal relationship between the points in the same line along the direction of width, height and time. However, only replacing the two CCA-2Ds in the RCCA-2D with our CCA-3Ds cannot model the spatiotemporal context in videos. Therefore, we further duplicate the CCA-3D with a recurrent mechanism to transmit the relation between the points in a line to a plane and finally to the whole spatiotemporal space. To make the RCCA-3D adaptive for action recognition, we propose a novel recurrent structure rather than directly extending the original 2D structure to 3D. In the experiments, we make a thorough analysis of different structures of RCCA-3D, verifying the proposed structure is more suitable for action recognition. We also compare our RCCA-3D with the non-local attention, showing that the RCCA-3D requires 25% fewer parameters and 30% fewer FLOPs with even higher accuracy. Finally, equipped with our RCCA-3D, 3 networks achieve better and leading performance on 5 RGB-based and skeleton-based datasets.

Unsupervised Video-Based Action Recognition With Imagining Motion and Perceiving Appearance

Video-based action recognition is a challenging task, which demands carefully considering the temporal property of videos in addition to the appearance attributes. Particularly, the temporal domain of raw videos usually contains significantly more redundant or irrelevant information than still images. For that, this paper proposes an unsupervised video-based action recognition approach with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">imagining motion</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">perceiving appearance</i> , called IMPA, by comprehensively learning the spatio-temporal characteristics inherited in videos, with a particular emphasis on the moving object for action recognition. Specifically, a self-supervised <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Motion Extracting Block</i> (MEB) is designed to extract the principal motion features by focusing on the large movement of the moving object, based on the observation that humans can infer complete motion trajectories from partial moving objects. To further take the indispensable appearance attribute in videos into account, an unsupervised <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Appearance Learning Block</i> (ALB) is developed to perceive the static appearance, thus in combination with the MEB to recognize actions. Extensive validation experiments and ablation studies on multiple datasets demonstrate that our proposed IMPA approach obtains superior performance and surpasses other classical and state-of-the-art unsupervised action recognition methods.

Global spatio-temporal synergistic topology learning for skeleton-based action recognition

Compared to RGB video-based action recognition, skeleton-based action recognition algorithm has attracted much more attention due to being more lightweight, better generalization and robustness. The extraction of temporal and spatial features is a crucial factor for skeleton-based action recognition. However, existing feature extraction methods suffer from two limitations: (1) the isolated extraction of temporal and spatial feature cannot capture temporal feature connections among non-adjacent joints and (2) convolution-limited perceptual fields cannot capture global temporal features of joints effectively. In this work, we propose a global spatio-temporal synergistic feature learning module (GSTL), which generates global spatio-temporal synergistic topology of joints by spatio-temporal feature fusion. By further combining the GSTL with a temporal modeling unit, we develop a powerful global spatio-temporal synergistic topology learning network (GSTLN), and it achieves competitive performance with fewer parameters on three challenge datasets: NTU RGB + D, NTU RGB + D 120, and NW-UCLA.

DeepActsNet: A deep ensemble framework combining features from face, hands, and body for action recognition

Human action recognition from videos has gained substantial focus due to its wide applications in the field of video understanding. Most of the existing approaches extract human skeleton data from videos to encode actions because of the invariance nature of the skeleton information with respect to lightning conditions and background changes. Despite their success in achieving high recognition accuracy, methods based on limited body joints fail to capture the nuances of subtle body parts which are highly relevant for discriminating similar actions. In this paper, we overcome this limitation by presenting a holistic framework for combining spatial and motion features from the body, face, and hands to develop a novel data representation termed “Deep Actions Stamps (DeepActs)” for video-based action recognition. Compared to the skeleton sequences based on limited body joints, DeepActs encode more effective spatio-temporal features that provide robustness against pose estimation noises and improve action recognition accuracy. We also present “DeepActsNet”, a deep learning based ensemble model which learns convolutional and structural features from Deep Action Stamps for highly accurate action recognition. Experiments on three challenging action recognition datasets (NTU60, NTU120, and SYSU) show that the proposed model produces significant improvements in the action recognition accuracy with less computational cost compared to the state-of-the-art methods.

Fall-Perceived Action Recognition of Persons With Neurological Disorders Using Semantic Supervision

Frequent uncertain falls is one of the common cause of injury among elderly adults and persons suffering from the neurological disorder. It will be costlier to go through <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$24\times 7$ </tex-math></inline-formula> medical monitoring if we monitor a person suffering from the early stage of the neurological disorder. An “uncertain” action classification model can be a less costly and easily scalable. It can help to regularly monitor a person suffering from neurological declines and how frequent it relapse. In this article, we propose a video-based action recognition with fall detection architecture, FallNet, which learns the features of uncertain actions related to day-to-day activities. FallNet first incorporates semantic supervision using the per-class weight of uncertain action through class-wise weighted focal loss. It addresses both the class imbalance problem and the weak interclass separability issue. We design a joint training model to train the overall architecture efficiently in an end-to-end manner. We utilize benchmark data sets, OOPS, HMDB51, and Kinetics-600, for experimentation that has less falling action videos. Therefore, we have collected videos to create a data set, denoted by FallAction, that consists of different 15 falling action classes with an average of 100 videos per class. The proposed network gain an accuracy of 13.2% in OOPs, 2% in HMDB51, and 0.2% in Kinetics-600 data set.

What we see is what we do: a practical Peripheral Vision-Based HMM framework for gaze-enhanced recognition of actions in a medical procedural task

Deep learning models have shown remarkable performances in egocentric video-based action recognition (EAR), but rely heavily on a large quantity of training data. In specific applications with only limited data available, eye movement data may provide additional valuable sensory information to achieve accurate classification performances. However, little is known about the effectiveness of gaze data as a modality for egocentric action recognition. We, therefore, propose the new Peripheral Vision-Based HMM (PVHMM) classification framework, which utilizes context-rich and object-related gaze features for the detection of human action sequences. Gaze information is quantified using two features, the object-of-interest hit and the object–gaze distance, and human action recognition is achieved by employing a hidden Markov model. The classification performance of the framework is tested and validated on a safety-critical medical device handling task sequence involving seven distinct action classes, using 43 mobile eye tracking recordings. The robustness of the approach is evaluated using the addition of Gaussian noise. Finally, the results are then compared to the performance of a VGG-16 model. The gaze-enhanced PVHMM achieves high classification performances in the investigated medical procedure task, surpassing the purely image-based classification model. Consequently, this gaze-enhanced EAR approach shows the potential for the implementation in action sequence-dependent real-world applications, such as surgical training, performance assessment, or medical procedural tasks.

Self-supervised Video-based Action Recognition With Disturbances.

Self-supervised video-based action recognition is a challenging task, which needs to extract the principal information characterizing the action from content-diversified videos over large unlabeled datasets. However, most existing methods choose to exploit the natural spatio-temporal properties of video to obtain effective action representations from a visual perspective, while ignoring the exploration of the semantic that is closer to human cognition. For that, a self-supervised Video-based Action Recognition method with Disturbances called VARD, which extracts the principal information of the action in terms of the visual and semantic, is proposed. Specifically, according to cognitive neuroscience research, the recognition ability of humans is activated by visual and semantic attributes. An intuitive impression is that minor changes of the actor or scene in video do not affect one person's recognition of the action. On the other hand, different humans always make consistent opinions when they recognize the same action video. In other words, for an action video, the necessary information that remains constant despite the disturbances in the visual video or the semantic encoding process is sufficient to represent the action. Therefore, to learn such information, we construct a positive clip/embedding for each action video. Compared to the original video clip/embedding, the positive clip/embedding is disturbed visually/semantically by Video Disturbance and Embedding Disturbance. Our objective is to pull the positive closer to the original clip/embedding in the latent space. In this way, the network is driven to focus on the principal information of the action while the impact of sophisticated details and inconsequential variations is weakened. It is worthwhile to mention that the proposed VARD does not require optical flow, negative samples, and pretext tasks. Extensive experiments conducted on the UCF101 and HMDB51 datasets demonstrate that the proposed VARD effectively improves the strong baseline and outperforms multiple classical and advanced self-supervised action recognition methods.

Neighbor-Guided Consistent and Contrastive Learning for Semi-Supervised Action Recognition.

Semi-supervised learning has been well established in the area of image classification but remains to be explored in video-based action recognition. FixMatch is a state-of-the-art semi-supervised method for image classification, but it does not work well when transferred directly to the video domain since it only utilizes the single RGB modality, which contains insufficient motion information. Moreover, it only leverages highly-confident pseudo-labels to explore consistency between strongly-augmented and weakly-augmented samples, resulting in limited supervised signals, long training time, and insufficient feature discriminability. To address the above issues, we propose neighbor-guided consistent and contrastive learning (NCCL), which takes both RGB and temporal gradient (TG) as input and is based on the teacher-student framework. Due to the limitation of labelled samples, we first incorporate neighbors information as a self-supervised signal to explore the consistent property, which compensates for the lack of supervised signals and the shortcoming of long training time of FixMatch. To learn more discriminative feature representations, we further propose a novel neighbor-guided category-level contrastive learning term to minimize the intra-class distance and enlarge the inter-class distance. We conduct extensive experiments on four datasets to validate the effectiveness. Compared with the state-of-the-art methods, our proposed NCCL achieves superior performance with much lower computational cost.

Deep learning-based activity recognition and fine motor identification using 2D skeletons of cynomolgus monkeys.

Video-based action recognition is becoming a vital tool in clinical research and neuroscientific study for disorder detection and prediction. However, action recognition currently used in non-human primate (NHP) research relies heavily on intense manual labor and lacks standardized assessment. In this work, we established two standard benchmark datasets of NHPs in the laboratory: MonkeyinLab (MiL), which includes 13 categories of actions and postures, and MiL2D, which includes sequences of two-dimensional (2D) skeleton features. Furthermore, based on recent methodological advances in deep learning and skeleton visualization, we introduced the MonkeyMonitorKit (MonKit) toolbox for automatic action recognition, posture estimation, and identification of fine motor activity in monkeys. Using the datasets and MonKit, we evaluated the daily behaviors of wild-type cynomolgus monkeys within their home cages and experimental environments and compared these observations with the behaviors exhibited by cynomolgus monkeys possessing mutations in the MECP2 gene as a disease model of Rett syndrome (RTT). MonKit was used to assess motor function, stereotyped behaviors, and depressive phenotypes, with the outcomes compared with human manual detection. MonKit established consistent criteria for identifying behavior in NHPs with high accuracy and efficiency, thus providing a novel and comprehensive tool for assessing phenotypic behavior in monkeys.

Action recognition based on discrete cosine transform by optical pixel-wise encoding

The framework provides a novel pipeline for action recognition. The action recognition task classifies the action label of the scene. High-speed cameras are commonly used to generate high frame-rate videos for capturing sufficient motion information. However, the data volume would be the bottleneck of the system. With the insight that the discrete cosine transform (DCT) of video signals reveals the motion information remarkably, instead of obtaining video data as with traditional cameras, the proposed method directly captures a DCT spectrum of video in a single shot through optical pixel-wise encoding. Considering that video signals are sparsely distributed in the DCT domain, a learning-based frequency selector is designed for pruning the trivial frequency channels of the spectrum. An opto-electronic neural network is designed for action recognition from a single coded spectrum. The optical encoder generates the DCT spectrum, and the rest of the network jointly optimizes the frequency selector and classification model simultaneously. Compared to conventional video-based action recognition methods, the proposed method achieves higher accuracy with less data, less communication bandwidth, and less computational burden. Both simulations and experiments demonstrate that the proposed method has superior action recognition performance. To the best of our knowledge, this is the first work investigating action recognition in the DCT domain.

Adaptive spatiotemporal graph convolutional network with intermediate aggregation of multi-stream skeleton features for action recognition

Video-based action recognition is a challenging problem due to the rapid and uncertain changes in human actions. Recent studies show that incorporating video and human body skeleton helps improve action recognition performance. These methods generally use graph convolutional networks (GCNs) to extract structural features of the human body joints from skeleton data. Yet, most GCN-based methods have some limitations in skeleton-based action recognition. (1) The graph structure of the human body joints is time-invariant, making it difficult to represent the changing relationship between joints across actions. (2) Methods relying on single-stream models only utilize limited information of skeleton data, such as joints or bones, and fail to consider coherent features of movements. (3) Methods relying on multi-stream models have considerable parameters and are inefficient for real-life applications. To address these problems, we propose an adaptive spatiotemporal graph convolutional network with intermediate aggregation of multi-stream skeleton features for action recognition. First, our method learns an adaptive graph structure representing the changing relationship between joints. Secondly, we facilitate a multi-stream model to extract various features from the skeleton, including joint-stream, bone-stream, and motion-stream. Moreover, an intermediate aggregation strategy is employed to aggregate these features and to reduce the parameters of this model. The proposed method has been validated on various benchmarks and a real-world abnormal action dataset. Extensive experimental results show that our method achieves excellent performance in skeleton-based action recognition.