Efficient Auxiliary Tasks for Skeleton-Based Action Recognition

Skeleton-based action recognition is a prominent research area focused on developing specialized graph convolutional networks (GCNs) to model 3D skeleton data. To enhance the capture of spatial and temporal dependencies in GCNs for extracting discriminative features, we introduce two auxiliary tasks: a Spatial-Temporal Masked Reconstruction (STMR) mechanism and a Skeleton Prototype Contrastive (SPC) learning strategy. The STMR mechanism leverages the spatial structure and temporal continuity of skeleton joints to reconstruct masked joints, enabling GCNs to identify more valuable patterns and semantics for action recognition. Meanwhile, the SPC learning strategy identifies characteristic skeleton features for each action and learns discriminative skeleton representations. Extensive experiments on the NTU RGB+D and NW-UCLA datasets demonstrate that the proposed method significantly improves action recognition performance.