Monocular Weakly-Supervised Camera-Relative 3D Human Pose Estimation

This paper presents a 3D human pose estimation framework based on Deep Neural Networks (DNNs). It builds upon existing weakly-supervised methods that predict 2D-3D correspondences and improves them by introducing a geometrical-alignment pre-processing step and a 3D skeleton-refinement post-processing step. The geometrical-alignment pre-processing step is applied on the ground-truth 3D human poses and transforms them, in order to enable the utilized 2D-to-3D skeleton mapping DNN to be efficiently trained in a weakly-supervised manner. The 3D skeleton-refinement post-processing step acts on the DNN outputs and enables the proposed 3D human pose estimation framework to predict the camera-relative 3D human poses. Experiments on the widely used public showed that the proposed framework managed to predict camera-relative 3D human poses with increased accuracy.