Learning to Exploit Stability for 3D Scene Parsing