See and Think: Disentangling Semantic Scene Completion