SkeletonPose: Exploiting human skeleton constraint for 3D human pose estimation

Abstract

We present SkeletonPose, a novel method to improve the accuracy of 3D human pose estimation via human skeleton constraints. In contrast to other state-of-the-art 3D pose estimation approaches, which use deep convolutional networks to regress the human pose, the proposed method employs a combination of data-driven and calculation methods. Our approach explicitly exploits the skeleton length prior during testing which can decrease the predicted human skeleton length error. As a result, the predicted 3D human pose error deceases accordingly. First, the proposed approach uses deep convolutional networks to regress the z-coordinate of the root joint. Second, the predicted outputs of the networks are used to calculate the 3D human pose according to the skeleton length invariance constraint. This combined method increases the accuracy of pose estimation because the skeleton length prior restricts the human pose space. Moreover, to eliminate the ambiguity of the z-coordinate difference between two connected joints, we propose a step-wise refinement to reduce the adverse effect caused by inaccurate predictions. Thorough evaluations were performed on three public databases; the Human3.6M dataset, the Human Eva-I database and the MPI-INF-3DHP dataset. Comparisons illustrate that SkeletonPose achieved better performance with respect to other state-of-the-art pose estimation approaches. The code and data are available at https://github.com/XTU-PR-LAB/skeleton-pose.