3D human pose estimation is a challenging but important research topic with abundant applications. As for discriminative human pose estimation, the main goal is to learn a nonlinear mapping from image descriptors to 3D human pose configurations, which is difficult due to the high-dimensionality of h