Learning Viewpoint Agnostic Visual Representations by Recovering