AtlasNet: A Papier-M\\ᵃch\\'e Approach to Learning 3D Surface Generation

We introduce a method for learning to generate the surface of 3D shapes. Our represents a 3D shape as a collection of parametric surface elements, in contrast to methods generating voxel grids or point clouds, naturally a surface representation of the shape. Beyond its novelty, our new shape framework, AtlasNet, comes with significant advantages, such as precision and generalization capabilities, and the possibility to a shape of arbitrary resolution without memory issues. We demonstrate benefits and compare to strong baselines on the ShapeNet benchmark for applications: (i) auto-encoding shapes, and (ii) single-view reconstruction a still image. We also provide results showing its potential for other, such as morphing, parametrization, super-resolution, matching, co-segmentation.