In order for the deep learning models to truly understand the 2D images for 3D geometry recovery, we argue that singleview reconstruction should be learned in a part-aware and weakly supervised manner. Such models lead to ...