Learning Representation for Scene Understanding: Epitomes, CRFs, and CNNs