Dynamic partitioning graph convolutional network for skeleton-based action recognition

Abstract

In skeleton-based action recognition, most state-of-the-art models are based on graph convolutional networks (GCNs), and propose various graph topologies to capture the inherent relationships between joints more effectively. However, existing GCN-based methods are constrained by their reliance on independently proposed graph topologies, and often overlook the potential benefits of incorporating various graphs. To address this limitation, we propose a dynamic partitioning GCN (DPGCN) that can capture the dynamic dependencies of the skeletal structure and learn the relationships among subgraphs. The DPGCN, inspired by dynamic convolution, includes a dynamic partitioning graph convolution (DP-GC) designed to extract features from subgraphs and predict dynamically adjusted weights. DP-GC assigns predicted weights to multiple kernels and combines them, allowing the model to focus on the structural patterns. Our proposed DPGCN outperforms or achieves a performance comparable to state-of-the-art methods on three benchmark datasets: NTU RGB+D, NTU RGB + D 120, and NW-UCLA.