Using deep neural networks, supervised 3D reconstruction has made significant progress. However, large-scale annotations of 2D/3D data are necessary for this performance boost [4]. How to effectively represent 3D data to feed deep networks remains a challenge in 3D deep learning. Volumetricor point cloud representations have been used inrecent works, but these methods have a number of drawbacks, including computational complexity, unorganized data, and a lack of finer geometry