3D Human Pose Estimation with Bone Symmetry and Motion Locality

The bone symmetry and motion locality are two important priors for 3D human pose estimation. For bone symmetry, symmetric bones typically have equal length and higher motion correlation. In terms of motion locality, the motion of the human body conforms to its structure and is temporally continuous. These two priors align with human intuition, but are challenging for Transformer to directly abstract them from data. We propose novel modules, Bone Symmetry Temporal Transformer (BSTT) and Spatial-Temporal Locality Fusion (STLF), to utilize these two priors, and design a new Transformer-based BSMLFormer model. Specifically, BSTT considers the temporal sequences of two symmetric bones as a single new sequence and calculates its global relationships by self-attention mechanism. STLF performs local computations by GCN and convolutions to capture the local dependencies in adjacent joints and adjacent parts. The local-global alternating structure of BSMLFormer allows for consideration of the local information during global calculations. More remarkably, BSMLFormer achieves the state-of-the-art performance with 39.0mm MPJPE on Human3.6M dataset.