Action recognition from arbitrary views using 3D-key-pose set