Hierarchical Spatial-Temporal Masked Contrast for Skeleton Action Recognition

In the field of 3-D action recognition, self-supervised learning has shown promising results but remains a challenging task. Previous approaches to motion modeling often relied on selecting features solely from the temporal or spatial domain, which limited the extraction of higher-level semantic information. Additionally, traditional one-to-one approaches in multilevel comparative learning overlooked the relationships between different levels, hindering the learning representation of the model. To address these issues, we propose the hierarchical spatial-temporal masked network (HSTM) for learning 3-D action representations. HSTM introduces a novel masking method that operates simultaneously in both the temporal and spatial dimensions. This approach leverages semantic relevance to identify meaningful regions in time and space, guiding the masking process based on semantic richness. This guidance is crucial for learning useful feature representations effectively. Furthermore, to enhance the learning of potential features, we introduce cross-level distillation (CLD) to extend the comparative learning approach. By training the model with two types of losses simultaneously, each level of the multilevel comparative learning process can be guided by levels rich in semantic information. This allows for more effective supervision of comparative learning, leading to improved performance. Extensive experiments conducted on the NTU-60, NTU-120, and PKU-MMD datasets demonstrate the effectiveness of our proposed framework. The learned action representations exhibit strong transferability and achieve state-of-the-art results.