Balance feature transferring GCN for 3D human pose estimation
Yangfei ZHAO, Renbo LIU, Weifeng Zhang, Yong NIU, Shuo Wang, Pei LV, Mingliang Xu · Scientia Sinica Informationis · 2025
In the field of 3D human pose estimation, existing graph convolutional network (GCN) methods commonly face the challenge of local over-smoothing. This issue arises from the scale differences in node receptive fields, leading to imbalanced feature smoothness evolution among joints, which severely hinders the effective extraction of full-body pose features. To address this, we propose a lightweight Balance GCN, which mitigates local over-smoothing through a balanced feature transferring mechanism. First, we construct a feature transferring balance metric based on node eccentricity to quantify the disparity in graph convolution operations required for nodes to perceive global features. Next, we design an apart transfer module that optimizes the transfer matrix under feature transferring balance constraints to guide inter-node feature exchange. Finally, the Balance GCN is integrated. Experiments on the Human3.6M dataset demonstrate that, compared to the Vanilla GCN baseline, Balance GCN significantly reduces MPJPE to 43.4 mm, achieving a 61.6% error reduction without increasing parameters or computational cost. Additionally, the apart transfer module exhibits strong generalizability, improving accuracy by an average of 1.59 mm across 9 mainstream methods.