Difference-attention graph convolutional network for skeleton-based gesture recognition

Graph Convolutional Networks (GCNs) have been widely applied to skeleton-based gesture recognition tasks and have achieved remarkable performance. The currently proposed dynamic and topology-non-shared graph convolutional networks outperform conventional graph convolutional networks. However, most existing studies rely solely on simple pairwise subtraction between nodes to construct dynamic adjacency matrices, which limits the in-depth exploration of potential relationships between nodes. And some models have the problem of large number of parameters, which leads to slow inference speed. To address these issues, we propose a Difference-Attention Graph Convolutional Network (DA-GCN), which is also a dynamic and non-shared graph convolutional network. First, the skeleton data is used to extract high-level spatiotemporal features. Then, the proposed Difference Relation Unit (DRU) and Attention Relation Unit (ARU) dynamically model the topological relationships in channel and temporal dimensions. These are integrated with the predefined adjacency matrix to construct a dynamic, non-shared adjacency matrix. Finally, based on the newly constructed adjacency matrix, the spatiotemporal features of skeletal joints are propagated between both neighboring and non-neighboring nodes. We conduct extensive experiments on three widely used benchmark datasets—SHREC’17 Track, DHG-14/28 and FPHA. The results demonstrate that our proposed method achieves competitive recognition performance across all three benchmark datasets. Meanwhile, it also exhibits superior recognition efficiency in terms of model parameters, computational cost, and inference time. Our code is publicly available on a code hosting platform. Our project website is available at https://github.com/WYDHWYDH/DA-GCN-Gesture .