GM-HRNet: Human Pose Estimation Based on Global Modeling

Convolutional neural networks have been widely used for human pose estimation tasks, but with some issues. It is limited to local receptive fields and it is difficult to capture global information. To address this problem, we propose the GM-HRNet network. The network aggregates the multi-stage feature information of HRNet, and makes full use of the criss-cross attention and channel attention to obtain the context information in the space and channel dimensions, and realizes the modeling of the global relationship, so as to effectively locate the keypoints of the human body. In this paper, HRNet is used as a benchmark to conduct experiments on MPII dataset. Experimental results show that the accuracy of GM-HRNet is 1.3 percent higher than that of HRNet under the same experimental conditions, which validates the effectiveness of the GM-HRNet model.