Learning shape spaces of dressed 3D human models in motion