Incorporating Part Appearances Symmetry for Human Pose Estimation

Abstract

Tree models for human pose estimation have been prevailed in the last decade, which are effective in human pose estimation. This paper aims to incorporate the appearance symmetry of human limb parts into tree model and address the problem of the wrong detection of human limbs. For a pair of symmetrical limbs, such as for legs and arms, their appearances are similar that can use a distance to represent. In our approach, within such a pictorial structure framework and human tree model, we decompose a limb estimation problem into a judgment of finding the appearance of symmetrical limb parts. This suggests that we need to find the possible estimation of one pose and then eliminate some wrong estimation poses, which their appearances of parts are not symmetrical. In a series of experiment in different methods, the results demonstrate that our methods improve performance for pose estimation on the Parse dataset, LSP dataset and Sport dataset.