Beyond Grids: Learning Graph Representations for Visual Recognition