An efficient self-attention network for skeleton-based action recognition

There has been significant progress in skeleton-based action recognition. Human skeleton can be naturally structured into graph, so graph convolution networks have become the most popular method in this task. Most of these state-of-the-art methods optimized the structure of human skeleton graph to obtain better performance. Based on these advanced algorithms, a simple but strong network is proposed with three major contributions. Firstly, inspired by some adaptive graph convolution networks and non-local blocks, some kinds of self-attention modules are designed to exploit spatial and temporal dependencies and dynamically optimize the graph structure. Secondly, a light but efficient architecture of network is designed for skeleton-based action recognition. Moreover, a trick is proposed to enrich the skeleton data with bones connection information and make obvious improvement to the performance. The method achieves 90.5% accuracy on cross-subjects setting (NTU60), with 0.89M parameters and 0.32 GMACs of computation cost. This work is expected to inspire new ideas for the field.