Learning 3D object-centric representation through prediction