Whole-body 2D human pose estimation based on human keypoints distribution constraint and adaptive Gaussian factor

In this paper, we propose a method for whole-body 2D human pose estimation on the basis of the human key point distribution constraint and adaptive gaussian factor by studying the positions of different body part human body keypoints and generating adaptive gaussian heatmap. Unlike traditional pose estimation, whole-body 2D human pose estimation requires the location of keypoints of human body parts as well as keypoints of the face, hands and feet. Owing to the physical structure limitations of the human body, hands and face and the scale differences of different body parts, the results of whole-body 2D human pose estimation methods are often affected. To solve this problem, we propose a new method based on the human key point distribution to learn the relationships between keypoints of different body parts, and generate heatmaps using adaptive gaussian factor to handle scale and density variations. Moreover, we propose the use of different reference points for position constraints according to the distribution relationship of different body part keypoints, and generate adaptive gaussian heatmap to improve the local and overall 2D whole-body pose estimation accuracy at the same time. On the COCO-WholeBody dataset, our method achieved 77.8% whole-body AP and 80.3% AR, which are 16.8% and 6.2% higher than the existing SOTAs methods respectively; on the Halpe-FullBody dataset, the whole-body AP reached 84.0%, verifying the robustness of the method.