Human Pose Estimation Based On Relevance Constrained Learning

2D human pose estimation is a research highlight in the current area of computer vision,which aims at detecting and locating the whole keypoints appearing in the images,such as knees and ankles.Recently,although deep learning methods make progress on human pose estimation,most of work focuses on the detection of keypoints while ignores the relevance of them.Due to physically global connection of human body,this article proposes to learn the feature of human pose estimation from the view of relevance of ketpoints.(1)propose a human pose estimation network based on expectation maximization(EMpose Net).Human pose estimation task can be regarded as a multi-task learning inherently with regarding each task as a different keypoint detection.Therefore,sharing feature information among weak related tasks will lead to a phenomenon called negative transfer.In order to avoid the influence of negative transfer,this article divides all the related keypoints in human pose into several groups.With Bayesian analysis,the relevant keypoints are regarded as hidden variables,and a network structure named EMpose Net is proposed to learn and predict the position of keypoints for the human pose estimation.Compared with most of the current methods of feature learning,EMpose Net improves the feature expression ability of the network and prevents it converging to local extremum.Experiments carried out on COCO dataset and MPII dataset show that EMpose Net has good performance.(2)propose a human pose estimation network based on causal inference(CIpose Net).As a result of that there are potential dependence among keypoints,this article proposes to establish the relevant constraints among the keypoints via causal inference and designs the CIpose Net.Firstly,the probability graph of human pose is formed based on the physical structure constraints of the human body.Through causal inference,this article intervenes the probability graph of human pose based on backdoor adjustment to find the probability constraints of keypoints.The results of experiment on popular dataset show that it is superior for CIpose Net to model the dependence of keypoints.