Learning to Reconstruct Symmetric Shapes Using Planar Parameterization of 3D Surface

Abstract

Shape priors have been a game changer to achieve robust 3D reconstruction. Prior knowledge encoded in trained networks has proven to be effective in generating images. Based on a similar paradigm, various methods were proposed to generate 3D shape from images. To generate a voxel or point cloud representation of 3D shapes these methods required adding an extra dimension to the deep network, to handle 3D data. Unlike these methods, we try to reconstruct 3D shape from images by using a parameterized representation of the shape. For a 3D model, the information is mainly concentrated on the surface. We perform iterative parameterization of the surface to obtain a planar representation. This representation is encoded with surface information to generate 2D geometry images, which can be conveniently learned using traditional deep neural networks without additional overhead. We propose an efficient iterative planar parameterization to represent regions of high Gaussian curvature in geometry images. Our experiments demonstrate that the proposed network learns detailed features and is able to reconstruct geometrically accurate shapes from single image. Our code is available at https://github.com/hrdkjain/LearningSymmetricShapes.