Seeing the Objects Behind the Parts: Learning Compositional Models for Visual Recognition