Multi-Person Pose Estimation Based on Hierarchical Residual-Like Connections

Abstract

Recent methods of multi-person pose estimation focus on different aspects to increase the accuracy of keypoints localization. Although fusing the multi-scale feature maps to improve the recognition accuracy has achieved great results, there still have some space to promote. In this paper, we present two novel modules to enhance the multi-scale feature and increase the range of receptive fields by constructing hierarchical residual-like connections. First, the channel shuffle unit and Res2 block are combined to fuse the different level of features in pyramid feature maps, which prompts feature information communication. Second, a new residual block is built to fuse both spatial and channel-wise information within local receptive fields at each layer, and the residual block used in original basic network structure is replaced. The experiment have been evaluated on the COCO keypoint benchmark, which shows that our approach achieves better results than the other state-of-the-arts.