Improved Masking Strategy for Self-Supervised Skeleton-Based Action Recognition

Masked autoencoders have achieved great success in self-supervised skeleton-based action recognition. However, existing masking strategies fail to fully consider the topological connections among joints. Motion Intensity-based masking strategies have been proposed to reflect the information richness of individual joints, but they ignore the relationships between joints. Considering that the adjacency matrix can represent the topological connections among joints, we propose a masking strategy that integrates motion intensity and the adjacency matrix. This strategy enhances the model's ability to focus on informative motion regions. Based on this, we design a self-supervised skeletonbased action recognition framework using masked autoencoders, where specific joints are masked to predict their corresponding motion, significantly enhancing the model's ability to model the contextual motion and topological relationships of skeletal data. Experimental results on the NTU RGB + D, NTU RGB + D 120, and PKU-MMD datasets demonstrate that our proposed masking strategy significantly improves model performance and achieves state-of-the-art results.