Semantic Scene Completion through Multi-Level Feature Fusion