Learning Physical Graph Representations of Visual Scenes