Summary: This paper proposes a 3d super resolution method. It first projects the reconstructed 3D model to 6 low resolution orthographic depth maps. A mask network and a depth network are trained to up-sample the corresponding depth maps to high resolution ones, and a 3D model carving strategy is applied to produce high resolution reconstructed result. In their experiments, the reconstructed results are out perform the previous state-of-the-art pix2mesh algorithm by first do a low resolution reconstruction & then do super-resolution. Novelty: The insight of the paper is leverage the difficulty of super-resolution in 3D to the field of well studied image super-resolution, so that the learning could be much easier.