Predicting Animation Skeletons for 3D Articulated Models via Volumetric Nets