Body Partition Refinement Dynamic Graph Convolutional Network for Skeleton-based Action Recognition

Graph convolutional networks (GCNs) are the most commonly used methods recently and have greatly pro-moted research on skeleton-based action recognition. Aggregating features guided by a topology graph and generating adja-cency matrices with semantic edges in the topology graph are important for the task. However, building a simple and applica-ble graph is a challenging problem. In this work, we construct a novel body partition refinement dynamic graph convolutional network (BPR-DGCN) for skeleton-based action recognition, in which we propose a new body partition refinement dynamic graph convolutional module (BPR-DGCM) to efficiently extract representative features. Furthermore, by combining the traditional physical connection graph with our proposed body partition connection graph, we construct a body partition refinement dynamic graph convolutional unit (BPR-DGCU), which adopts a channel grouping strategy to learn a shared dynamic topology graph as a generic prior for all groups and refine it with chan-nel-specific body partition connection graph for different groups. Our BPR-DGCN outperforms state-of-the-art methods on the NTU RGB+D 60 dataset.