Abstract
This paper presents a novel anatomy-aware pre-training method for accurate 3D human pose estimation, named APTPose. We propose a Hierarchical Masked Pose Modeling (HMPM) subtask that decouples the body skeleton into several distinct body components for hierarchical modeling. It surpasses the limitations of earlier joint coordinate masking techniques by better capturing the dependencies of the human skeletal structure. Unlike previous methods focusing on 2D pose reconstruction in their pre-training task, we leverage a large number of 3D pseudo labels from existing datasets for pre-training. This allows us to better model the skeletal system in 3D space and improve the accuracy and robustness of 3D human pose estimation. Additionally, we introduce a geometric loss into the optimization process to boost correlations within the human skeleton. Experimental results show its superior robustness and generalization capabilities across challenging benchmarks, offering a favorable balance between accuracy and computational complexity, thus making it an appealing option for practical applications.