Vision Transformer, which performs well in various vision tasks, encounters a bottleneck in skeleton-based action recognition and falls short of advanced GCN-based methods. The root cause is that the current skeleton transformer depends on the self-attention mechanism of the complete channel of the global joint, ignoring the highly discriminative differential correlation within the channel, so it is challenging to learn the expression of the multivariate topology dynamically. To tackle this, we present Skeleton MixFormer, an innovative spatio-temporal architecture to effectively represent the physical correlations and temporal interactivity of the compact skeleton data. Two essential components make up the proposed framework: 1) Spatial MixFormer. The channel-grouping and mix-attention are utilized to calculate the dynamic multivariate topological relationships. Compared with the full-channel self-attention method, Spatial MixFormer better highlights the channel groups' discriminative differences and the joint adjacency's interpretable learning. 2) Temporal MixFormer, which consists of Multiscale Convolution, Temporal Transformer and Sequential Holding Module. The multivariate temporal models ensure the richness of global difference expression and realize the discrimination of crucial intervals in the sequence, thereby enabling more effective learning of long and short-term dependencies in actions. Our Skeleton MixFormer demonstrates state-of-the-art (SOTA) performance across seven different settings on four standard datasets, namely NTU-60, NTU-120, NW-UCLA, and UAV-Human. Related code will be available on https://github.com/ElricXin/Skeleton-MixFormer.

Gesture recognition is an important direction in computer vision research. Information from the hands is crucial in this task. However, current methods consistently achieve attention on hand regions based on estimated keypoints, which will significantly increase both time and complexity, and may lose position information of the hand due to wrong keypoint estimations. Moreover, for dynamic gesture recognition, it is not enough to consider only the attention in the spatial dimension. This paper proposes a multi-scale attention 3D convolutional network for gesture recognition, with a fusion of multimodal data. The proposed network achieves attention mechanisms both locally and globally. The local attention leverages the hand information extracted by the hand detector to focus on the hand region, and reduces the interference of gesture-irrelevant factors. Global attention is achieved in both the human-posture context and the channel context through a dual spatiotemporal attention module. Furthermore, to make full use of the differences between different modalities of data, we designed a multimodal fusion scheme to fuse the features of RGB and depth data. The proposed method is evaluated using the Chalearn LAP Isolated Gesture Dataset and the Briareo Dataset. Experiments on these two datasets prove the effectiveness of our network and show it outperforms many state-of-the-art methods.

Human action recognition (HAR), which aims to recognize and understand individual actions and intentions, has rapidly become a research hotspot in computer vision. Compared with other data modalities, skeleton data offers more efficient node semantics and more coherent spatio-temporal motion patterns, effectively reducing the impact of lighting and background changes. In recent years, many researchers have focused on skeleton-based action recognition methods and have made significant progress. However, we believe that the current skeleton-based action recognition methods still face three major challenges: 1) reducing reliance on expensive labeled data while maintaining model performance; 2) enabling the model to understand and recognize new behavior classes with a limited number of samples; and 3) addressing the challenges posed by the lack of skeleton information in single-modality spatio-temporal motion representation learning. Based on these challenges, we conduct a comprehensive review of the existing skeleton-based action recognition methods. Additionally, we provide an extensive review and analysis of publicly available action recognition datasets. This review aims to offer researchers a comprehensive perspective, stimulate more innovative ideas, and promote the application and breakthrough of skeleton action recognition in a wider range of computer vision tasks.

Skeleton-based action recognition is a crucial but challenging task in the application of engineering algorithms. However, due to the inaccurate estimation quality, certain joints that should theoretically lack dynamic information show irregular jitter, which affects the recognition accuracy. In this paper, we propose a Cascade Interaction Spatial-temporal transformer Network ( CI-STFormer ) to address abnormal joint jitter and multi-scale aggregation of features effectively. The network comprises three parts: (1) Spatial sparse-level optimization and interaction module. Relying on the topology map, the sparse-level interaction adjacency matrix is constructed through cascading interaction to mask the jitter non-discriminative joints, and then the transformer is applied to realize the sparse-level internal feature update. (2) Spatial primal-level fusion and interaction module. Multi-head self-attention is employed to enhance the original feature representation. Then, the sparse-level update features are cascaded to construct the scale-cascaded interaction adjacency matrix to achieve the balance of noise caused by the jitter and the fusion of different scale features. (3) Temporal domain scale-level TCN combines the multi-scale filter and the temporal self-attention channel interaction algorithm to extract different temporal features and the interaction and update of global temporal information. The experimental results show that the proposed method performs excellently on the four datasets of NTU RGB+D 60, NTU RGB+D 120, UAV-Human, and NW-UCLA and achieves state-of-the-art performance on the transformer-based track. Related code will be available on https://github.com/Xdu-Liu/CI-STFormer.git.

Graph convolution networks (GCNs) have achieved remarkable performance in skeleton-based action recognition by exploiting the adjacency topology of body representation. However, the adaptive strategy adopted by the previous methods to construct the adjacency matrix is not balanced between the performance and the computational cost. We assume this concept of Adaptive Trap, which can be replaced by multiple autonomous submodules, thereby simultaneously enhancing the dynamic joint representation and effectively reducing network resources. To effectuate the substitution of the adaptive model, we unveil two distinct strategies, both yielding comparable effects: (1) Optimization. Individuality and Commonality GCNs (IC-GCNs) is proposed to specifically optimize the construction method of the associativity adjacency matrix for adaptive processing. The uniqueness and co-occurrence between different joint points and frames in the skeleton topology are effectively captured through methodologies like preferential fusion of physical information, extreme compression of multi-dimensional channels, and simplification of self-attention mechanism. (2) Replacement. AutoLearning GCNs (AL-GCNs) is proposed to boldly remove popular adaptive modules and cleverly utilize human key points as motion compensation to provide dynamic correlation support. AL-GCNs construct a fully learnable group adjacency matrix in both spatial and temporal dimensions, resulting in an elegant and efficient GCN-based model. In addition, three effective tricks for skeletonbased action recognition (Skip-Block, Bayesian Weight Selection Algorithm, and Simplified Dimensional Attention) are exposed and analyzed in this paper. Finally, we employ the variable channel and grouping method to explore the hardware resource bound of the two proposed models. IC-GCN and AL-GCN exhibit impressive performance across NTU-RGB+D 60, NTU-RGB+D 120, NW-UCLA, and UAV-Human datasets, with an exceptional parameter-cost ratio. Code will be available on Adaptive-Pitfall.

In recent years, skeleton-based action recognition, graph convolutional networks, have achieved remarkable performance. In these existing works, the features of all nodes in the neighbor set are aggregated into the updated features of the root node, while these features are located in the same feature channel determined by the same $1\times 1$ convolution filter. This may not be optimal for capturing the features of spatial dimensions among adjacent vertices effectively. Besides, the effect of feature channels that are independent of the current action on the performance of the model is rarely investigated in existing methods. In this paper, we propose cross-channel graph convolutional networks for skeleton-based action recognition. The features fusion mechanism in our network is cross-channel, i.e, the updated feature of the root node is derived from different feature channels. Because different feature channels come from different $1 \times 1$ convolution filters, the cross-channel fusion mechanism significantly improves the ability of the model to capture local features among adjacent vertices. Moreover, by introducing a channel attention mechanism to our model, we suppress the influence of feature channels unrelated to action recognition on model performance, which improves the robustness of the model against the feature channels independent of the current action. Extensive experiments on the two large-scale datasets, NTU-RGB+D and Kinetics-Skeleton, demonstrate that the performance of our model exceeds the current mainstream methods.