Branch Information Correction Network for Human Pose Estimation

Qingzhan Ni, Chenxing Wang, Feipeng Da

Abstract: The main task of human keypoint detection is to detect the position of human bone joints in pictures or videos. In the branch-based network, key points are classified according to different properties, and each branch of the network is responsible for the prediction of a certain set of keypoints. Compared with the direct prediction of the network of all the key points, the advantage of this method is that it considers the structural constraints on the human body and the internal relationship between the key points. Based on the branch network, this paper studies the information-sharing relationship and the negative transfer relationship between branches. It proposes a new branch information correction network to make full use of the complementary information on branches. The experiment proves that the method proposed in this paper can further improve the accuracy of the keypoint prediction of the human body, and can correct some key points which are easily affected by the environment.