Learning Contextual Information For Holistic Scene Understanding