Advancing Multi-actor Graph Convolutions for Skeleton-Based Action Recognition

Human skeleton motion recognition, notable for its lightweight, interference-resistant, and resource-saving properties, plays a crucial role in human motion recognition and has found widespread applications. The common approach to capture motion features from human skeleton videos involves extracting skeleton features temporally or spatially using Graph Convolution Networks (GCN) or their improved variants. Nevertheless, existing extraction methods encounter two primary limitations: variability in the number of actors involved in an action and disconnected subgraphs representing multiple actors’ actions, resulting in a loss of inter-subgraph features. To overcome these challenges, we propose Human Mirror and Human Link strategies, which replicate diverse human data to fill and interlink multiple subgraphs. Empirically, our proposed methods applied to the NTU RGB+D 120 dataset significantly enhanced the performance of the base model MSG3D, demonstrating the effectiveness of our approach in handling multi-actor scenarios.