Two Stream Multi-Attention Graph Convolutional Network for Skeleton-Based Action Recognition

The skeleton-based action recognition has attracted much attention of researchers. The existing methods mostly introduce motion information into models by using multi-stream architecture, which leads to more parameters and FLOPs. In this paper, to resolve this problem, the proposed 2s-MAGCN (Two-Stream Multi-Attention Graph Convolutional Network) introduces motion information by applying the Motion Excitation attention module, which not only leads to less parameters and FLOPs by merging multi-stream into two-stream, but also improves the performance of the model. By proposing new strategies of pooling operations in attention modules, we get attention modules with better performance. It includes Spatial Excitation and Temporal Excitation, which are proposed to enhance the spatio-temporal expression ability of the model. On cross-subject benchmark and cross-view benchmark of NTU-RGB+D datasets, the proposed model achieves 88.60% and 97.16% accuracy respectively, and 35.62% accuracy on the Kinetics dataset. On both datasets, our method outperforms state-of-the-art methods.