Explore the vast range of titles available.

Recent methods for video action recognition have reached outstanding performances on existing benchmarks. However, they tend to leverage context such as scenes or objects instead of focusing on understanding the human action itself. For instance, a tennis field leads to the prediction playing tennis irrespectively of the actions performed in the video. In contrast, humans have a more complete understanding of actions and can recognize them without context. The best example of out-of-context actions are mimes, that people can typically recognize despite missing relevant objects and scenes. In this paper, we propose to benchmark action recognition methods in such absence of context and introduce a novel dataset, Mimetics, consisting of mimed actions for a subset of 50 classes from the Kinetics benchmark. Our experiments show that (a) state-of-the-art 3D convolutional neural networks obtain disappointing results on such videos, highlighting the lack of true understanding of the human actions and (b) models leveraging body language via human pose are less prone to context biases. In particular, we show that applying a shallow neural network with a single temporal convolution over body pose features transferred to the action recognition problem performs surprisingly well compared to 3D action recognition methods.

The number of categories for action recognition is growing rapidly and it has become increasingly hard to label sufficient training data for learning conventional models for all categories. Instead of collecting ever more data and labelling them exhaustively for all categories, an attractive alternative approach is “zero-shot learning” (ZSL). To that end, in this study we construct a mapping between visual features and a semantic descriptor of each action category, allowing new categories to be recognised in the absence of any visual training data. Existing ZSL studies focus primarily on still images, and attribute-based semantic representations. In this work, we explore word-vectors as the shared semantic space to embed videos and category labels for ZSL action recognition. This is a more challenging problem than existing ZSL of still images and/or attributes, because the mapping between video space-time features of actions and the semantic space is more complex and harder to learn for the purpose of generalising over any cross-category domain shift. To solve this generalisation problem in ZSL action recognition, we investigate a series of synergistic strategies to improve upon the standard ZSL pipeline. Most of these strategies are transductive in nature which means access to testing data in the training phase. First, we enhance significantly the semantic space mapping by proposing manifold-regularized regression and data augmentation strategies. Second, we evaluate two existing post processing strategies (transductive self-training and hubness correction), and show that they are complementary. We evaluate extensively our model on a wide range of human action datasets including HMDB51, UCF101, Olympic Sports and event datasets including CCV and TRECVID MED 13. The results demonstrate that our approach achieves the state-of-the-art zero-shot action recognition performance with a simple and efficient pipeline, and without supervised annotation of attributes. Finally, we present in-depth analysis into why and when zero-shot works, including demonstrating the ability to predict cross-category transferability in advance.

This paper presents the simultaneous utilization of video images and inertial signals that are captured at the same time via a video camera and a wearable inertial sensor within a fusion framework in order to achieve a more robust human action recognition compared to the situations when each sensing modality is used individually. The data captured by these sensors are turned into 3D video images and 2D inertial images that are then fed as inputs into a 3D convolutional neural network and a 2D convolutional neural network, respectively, for recognizing actions. Two types of fusion are considered—Decision-level fusion and feature-level fusion. Experiments are conducted using the publicly available dataset UTD-MHAD in which simultaneous video images and inertial signals are captured for a total of 27 actions. The results obtained indicate that both the decision-level and feature-level fusion approaches generate higher recognition accuracies compared to the approaches when each sensing modality is used individually. The highest accuracy of 95.6% is obtained for the decision-level fusion approach.

Over the past few decades, numerous studies have focused on identifying and recognizing human actions using machine learning and computer vision techniques. Video-based human action recognition (HAR) aims to detect actions from video sequences automatically. This can cover simple gestures to complex actions involving multiple people interacting with objects. Actions in team sports exhibit a different nature compared to other sports, since they tend to occur at a faster pace and involve more human-human interactions. As a result, research has typically not focused on the challenges of HAR in team sports. This paper comprehensively summarises HAR-related research and applications with specific focus on team sports such as football (soccer), basketball and Australian rules football. Key datasets used for HAR-related team sports research are explored. Finally, common challenges and future work are discussed, and possible research directions identified.

Existing public domain multi-modal datasets for human action recognition only include actions of interest that have already been segmented from action streams. These datasets cannot be used to study a more realistic action recognition scenario where actions of interest occur randomly and continuously among actions of non-interest or no actions. It is more challenging to recognize actions of interest in continuous action streams since the starts and ends of these actions are not known and need to be determined in an on-the-fly manner. Furthermore, there exists no public domain multi-modal dataset in which video and inertial data are captured simultaneously for continuous action streams. The main objective of this paper is to describe a dataset that is collected and made publicly available, named Continuous Multimodal Human Action Dataset (C-MHAD), in which video and inertial data stream are captured simultaneously in a continuous way. This dataset is then used in an example recognition technique and the results obtained indicate that the fusion of these two sensing modalities increases the F1 scores compared to using each sensing modality individually.

Understanding actions in videos remains a significant challenge in computer vision, which has been the subject of several pieces of research in the last decades. Convolutional neural networks (CNN) are a significant component of this topic and play a crucial role in the renown of Deep Learning. Inspired by the human vision system, CNN has been applied to visual data exploitation and has solved various challenges in various computer vision tasks and video/image analysis, including action recognition (AR). However, not long ago, along with the achievement of the transformer in natural language processing (NLP), it began to set new trends in vision tasks, which has created a discussion around whether the Vision Transformer models (ViT) will replace CNN in action recognition in video clips. This paper conducts this trending topic in detail, the study of CNN and Transformer for Action Recognition separately and a comparative study of the accuracy-complexity trade-off. Finally, based on the performance analysis’s outcome, the question of whether CNN or Vision Transformers will win the race will be discussed.

Recently, Transformer-based video recognition models have achieved state-of-the-art results on major video recognition benchmarks. However, their high inference cost significantly limits research speed and practical use. In video compression, methods considering small motions and residuals that are less informative and assigning short code lengths to them (e.g., MPEG4) have successfully reduced the redundancy of videos. Inspired by this idea, we propose Informative Patch Selection (IPS), which efficiently reduces the inference cost by excluding redundant patches from the input of the Transformer-based video model. The redundancy of each patch is calculated from motions and residuals obtained while decoding a compressed video. The proposed method is simple and effective in that it can dynamically reduce the inference cost depending on the input without any policy model or additional loss term. Extensive experiments on action recognition demonstrated that our method could significantly improve the trade-off between the accuracy and inference cost of the Transformer-based video model. Although the method does not require any policy model or additional loss term, its performance approaches that of existing methods that do require them.

In human action recognition, both spatio-temporal videos and skeleton features alone can achieve good recognition performance, however, how to combine these two modalities to achieve better performance is still a worthy research direction. In order to better combine the two modalities, we propose a novel Cross-Modal Transformer for human action recognition—CMF-Transformer, which effectively fuses two different modalities. In spatio-temporal modality, video frames are used as inputs and directional attention is used in the transformer to obtain the order of recognition between different spatio-temporal blocks. In skeleton joint modality, skeleton joints are used as inputs to explore more complete correlations in different skeleton joints by spatio-temporal cross-attention in the transformer. Subsequently, a multimodal collaborative recognition strategy is used to identify the respective features and connectivity features of two modalities separately, and then weight the identification results separately to synergistically identify target action by fusing the features under the two modalities. A series of experiments on three benchmark datasets demonstrate that the performance of CMF-Transformer in this paper outperforms most current state-of-the-art methods.

In recent years, skeleton-based human action recognition has garnered significant research attention, with proposed recognition or segmentation methods typically validated on large-scale coarse-grained action datasets. However, there remains a lack of research on the recognition of small-scale fine-grained human actions using deep learning methods, which have greater practical significance. To address this gap, we propose a novel approach based on heatmap-based pseudo videos and a unified, general model applicable to all modality datasets. Leveraging anthropometric kinematics as prior information, we extract common human motion features among datasets through an ad hoc pre-trained model. To overcome joint mismatch issues, we partition the human skeleton into five parts, a simple yet effective technique for information sharing. Our approach is evaluated on two datasets, including the public Nursing Activities and our self-built Tai Chi Action dataset. Results from linear evaluation protocol and fine-tuned evaluation demonstrate that our pre-trained model effectively captures common motion features among human actions and achieves steady and precise accuracy across all training settings, while mitigating network overfitting. Notably, our model outperforms state-of-the-art models in recognition accuracy when fusing joint and limb modality features along the channel dimension.

Regular exercise is a crucial aspect of daily life, as it enables individuals to stay physically active, lowers the likelihood of developing illnesses, and enhances life expectancy. The recognition of workout actions in video streams holds significant importance in computer vision research, as it aims to enhance exercise adherence, enable instant recognition, advance fitness tracking technologies, and optimize fitness routines. However, existing action datasets often lack diversity and specificity for workout actions, hindering the development of accurate recognition models. To address this gap, the Workout Action Video dataset (WAVd) has been introduced as a significant contribution. WAVd comprises a diverse collection of labeled workout action videos, meticulously curated to encompass various exercises performed by numerous individuals in different settings. This research proposes an innovative framework based on the Attention driven Residual Deep Convolutional-Gated Recurrent Unit (ResDC-GRU) network for workout action recognition in video streams. Unlike image-based action recognition, videos contain spatio-temporal information, making the task more complex and challenging. While substantial progress has been made in this area, challenges persist in detecting subtle and complex actions, handling occlusions, and managing the computational demands of deep learning approaches. The proposed ResDC-GRU Attention model demonstrated exceptional classification performance with 95.81% accuracy in classifying workout action videos and also outperformed various state-of-the-art models. The method also yielded 81.6%, 97.2%, 95.6%, and 93.2% accuracy on established benchmark datasets, namely HMDB51, Youtube Actions, UCF50, and UCF101, respectively, showcasing its superiority and robustness in action recognition. The findings suggest practical implications in real-world scenarios where precise video action recognition is paramount, addressing the persisting challenges in the field. The WAVd dataset serves as a catalyst for the development of more robust and effective fitness tracking systems and ultimately promotes healthier lifestyles through improved exercise monitoring and analysis.