Understanding the sources of error for 3D human pose estimation from monocular images and videos