Multi-scale Recalibration with Advanced Geometry Constraints for 3D Human Pose Estimation

Estimating 3D human poses from a single RGB image is a challenging task in computer vision due to depth ambiguity and lack of unconstrained 3D datasets. Poor exploiting of image cues will lead to inaccurate pose predictions. Previous works try to constrain illegal spatial relations and exploit 2D annotations of in-the-wild images as weak labels. In this paper, we propose to use multi-scale recalibration with stronger geometry constraints to regress 3D pose. The overall network is end-to-end which consists of a 2D pose estimation sub-network and a depth regression subnetwork. Firstly, by adding the residual multi-scale recalibration module to the depth regression sub-network, our approach works better on exploiting the structural information of human poses. Secondly, the advanced geometry constraints are introduced to restrict the human pose from the bone length and joint angle orientation, which shows a better result. Experimental results on the Human3.6m and in-the-wild MPII datasets show the effectiveness and robustness of our method.