Improving 3D Human Pose Estimation with Enhanced Body Feature Representation

Transformer-based methods have shown promising performance in 2D-to-3D human pose estimation. However, existing approaches often ignore the relationships between specific joints of the human body and the varying motion patterns across different body parts. This paper first introduces the Spatial Transformer Based on Human Anatomy module to enhance the spatial features of the human body. Additionally, a Cross-Joint Split and Recombination training strategy is designed within the temporal attention module to learn the motion patterns across different body parts. Finally, a Residual Gated Channel Transformation is introduced to foster collaboration and interaction between features. Experimental results on two benchmark datasets show that our algorithm effectively improves model performance with only a slight increase in computational cost.