Compositional Graph Convolutional Networks for 3D Human Pose Estimation

3D human pose estimation (HPE) from a single image can be decomposed into two subtasks, i.e., 2D HPE followed by 2D-to-3D pose lifting. Despite recent success in 2D HPE, 3D pose regression from 2D detections remains challenging due to the substantial depth ambiguity. Recently, graph convolutional networks (GCNs) have been exploited to model the relationships among body joints and demonstrate promising results. In this paper, we go one step further along this direction and propose a novel framework, termed Compositional GCN, for 3D HPE. It learns compositional relationships among body parts of different semantic levels and then exploits multilevel structural reasoning to reduce the depth uncertainty. Furthermore, we introduce a novel part-aware graph convolution. It not only disentangles self and neighbor transformations but also captures different relational patterns between each part and their respective neighbors. Experimental results demonstrate the effectiveness of the proposed approach.