Learning Scene Dynamics from Point Cloud Sequences