Spatial-Temporal Heatmap Masked Autoencoder for Skeleton-Based Action Recognition.

Skeleton representation learning offers substantial advantages for action recognition by encoding intricate motion details and spatial-temporal dependencies among joints. However, fully supervised approaches necessitate large amounts of annotated data, which are often labor-intensive and costly to acquire. In this work, we propose the Spatial-Temporal Heatmap Masked Autoencoder (STH-MAE), a novel self-supervised framework tailored for skeleton-based action recognition. Unlike coordinate-based methods, STH-MAE adopts heatmap volumes as its primary representation, mitigating noise inherent in pose estimation while capitalizing on advances in Vision Transformers. The framework constructs a spatial-temporal heatmap (STH) by aggregating 2D joint heatmaps across both spatial and temporal axes. This STH is partitioned into non-overlapping patches to facilitate local feature learning, with a masking strategy applied to randomly conceal portions of the input. During pre-training, a Vision Transformer-based autoencoder equipped with a lightweight prediction head reconstructs the masked regions, fostering the extraction of robust and transferable skeletal representations. Comprehensive experiments on the NTU RGB+D 60 and NTU RGB+D 120 benchmarks demonstrate the superiority of STH-MAE, achieving state-of-the-art performance under multiple evaluation protocols.