Hierarchical parallel multi-scale graph network for 3d human pose estimation

Graph convolutional network (GCN) is a widespread architecture for 2D-to-3D human pose estimation (HPE). Vanilla graph convolution is the key in GCN for body joints feature extraction by aggregating features from the first-order neighbors of each joint at single-scale. However, the nodes features updated in graph are gradually assimilated with the network deepens This makes it difficult to model the powerful feature representation for joints and has an adverse effect on resolving the uncertainty caused by occlusion or depth ambiguity. To address these problems, we propose a hierarchical parallel multi-scale graph convolutional network (HPM-GNet) for 3D HPE in this paper. The proposed network is composed by multi-scale sub-graph networks in parallel framework, guided by the geometric constraint of human body. Firstly, a well-designed weight masked graph convolutional (WMGConv) layer is used as a fundamental unit to construct the parallel multi-scale sub-graph convolutional network module (PMGCN) in HPM-GNet, which corporately captures the features of target nodes and neighbor nodes with weight assignment. Then, a cross-scale features exchange fusion block (CFEB) is designed to aggregate multi-scale features from local and global perspective with considering the geometric information from different parts of human body. Finally, we conduct experiments on two challenging 3D human pose benchmark datasets to evaluate the effectiveness of the proposed model. The experimental results demonstrate that our model achieves state-of-the-art performance.