Shape-Pose Ambiguity in Learning 3D Reconstruction from Images