Multibranch Attention Graph Convolutional Networks for 3-D Human Pose Estimation
Yanfang Yin, Ming Liu, Qigang Zhu, Shuaishuai Zhang, Naseer Ali Hussien, Yong Fan
Abstract
The human pose has a natural graphic structure. Therefore, in recent years, many studies on human pose estimation is based on Graph Convolutional Networks (GCN). However, the classical GCN method has a small receptive field and the transformation matrix is shared by all nodes, which limits their feature extraction ability. To address these issues, a novel multi-branch attention graph convolution (MultiBA_GConv) operation is proposed and a regression network model for human 3D position estimation based on the MultiBA_GConv operation is developed in this paper. Several different transformation matrices are used to extract the feature information contributing to the node itself, its neighbors, and other global nodes, and the corresponding attention mechanism is used to focus on these features in the MultiBA_GConv operation. The ablation experiments show that the multi-branch processing and multi-attention mechanism proposed in the MultiBA_GConv operation can greatly improve the feature extraction ability of graph convolution. The comparison experiments with other GCN methods show that MultiBA_GConv has the best feature extraction performance in all graph convolution operations. Based on the MultiBA_GConv, a 3D human pose regression network architecture is designed and the proposed method is evaluated on a challenging benchmark dataset for 3D human pose estimation. The experimental results of the comparison with the state of the art show that our method is highly competitive.