Is Really Correlation Information Represented Well in Self-Attention for Skeleton-based Action Recognition?

Transformer has shown significant advantages by various vision tasks. However, the lack of representation of correlation information about data properties makes it difficult to match the excellent results consistent with GCNs in skeleton-based action recognition. In this paper, we propose a Topology and Frames-guided Spatial-Temporal ConvFormer Network (TF-STCFormer), which is well suited for dynamically extracting topological and inter-frame uniqueness & co-occurrence information. Three essential components make up the proposed framework: (1) Grouped Physical-guided Spatial Transformer for focusing on learning essential spatial features and physical topology. (2) Global and Focal Temporal Transformer for promoting the relationship of different joints in consecutive frames and improving the representation of discriminative key-frames. (3) Grouped Dilation Temporal Convolution for connecting the intermediate output obtained by the previous transformers in the feature channels of different dilation. Experiments on four standard datasets (NTU RGB+D, NTU RGB+D 120, NW-UCLA, and UAV-Human) demonstrate that our approach prominently outperforms state-of-the-art methods on all benchmarks.