Skeleton-Based Action Recognition with Combined Part-Wise Topology Graph Convolutional Networks

Graph Convolutional Network (GCN) has achieved promising performance in skeleton-based action recognition by modeling skeleton sequences as spatio-temporal graphs. However, most existing methods only focus on the overall characteristics of the skeleton, thus lacking fine-grained exploration of human body parts semantics. In this paper, we propose a novel Combined Part-wise Topology Graph Convolutional Network (CPT-GCN), including SPT-GC, TPT-GC, and STPT-GC modules, to refine the spatio-temporal topology from the spatial, temporal, and spatio-temporal perspectives, respectively. Specifically, SPT-GC aggregates spatial features by combining global topology and partial correlations. TPT-GC combines the overall motion trend and the motion details of parts to capture temporal dynamics. STPT-GC establishes a spatio-temporal dependency, focusing on exploiting the implicit spatio-temporal information in motions. Ultimately, the effectiveness of CPT-GCN is demonstrated through experiments on two large-scale datasets: NTU RGB+D 60 and NTU RGB+D 120.