Parallel Multi-Branch Model with Multi-Resolution Interaction for Human Pose Estiation

• Published in: Journal of physics. Conference series Vol. 1865; no. 4; p. 42083
• Main Authors: Cai, Lingyi, Liu, Wei
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.04.2021
• Subjects: Activity recognition; Computer vision; Feature maps; Physics
• ISSN: 1742-6588; 1742-6596
• DOI: 10.1088/1742-6596/1865/4/042083

Abstract In recent years, human pose estimation has become a research hotspot in the field of computer vision, and has extensive application value in many fields, such as activity recognition [1], action detection, self-drving, etc. This paper tries a novel multi-stage network structure to complete the task of human pose estimation, and focus on improving the loss of spatial accuracy caused by repeated downsampling and upsampling. The network structure contains multiple resolution subnets. After the original resolution is learned in the first stage, a low-resolution branch will be generated in each subsequent stage, and these branches can learn feature representations in parallel. Through continuous multi-scale and multi-resolution fusion, necessary information interaction between subnets of different resolutions can be carried out to obtain richer feature maps and make the predicted key point heat maps more accurate. This paper uses COCO keypoint detection dataset to obtain the relevant results of pose estimation, which verifies the effectiveness of this network.