Primal-Dual Graph Attention Networks for 3D Human Pose Estimation

Graph convolution networks (GCNs) have shown good performance in 3D human pose estimation. However, GCNs are limited as they only focus on neighboring nodes for each node and have a fixed graph structure. To address these limitations, we propose a novel architecture called Primal-Dual Graph Attention Network. It consists of two parallel branches – the Primal branch, which applies feature-mask attention along with feature aggregation and attention across body-joints, and the Dual branch, which applies body-joint-mask attention and computes the correlation between feature dimensions. We sum the output of these two branches to get the estimated pose in 3D. This structure helps the network to generate improved feature representations and thus improves model performance. In this research, we use the public Human3.6M dataset.