AtlasNet

Why we included this project

AtlasNet is the reference implementation of a well-known CVPR 2018 method that turns either a single 2D image or a low-resolution point cloud into a full 3D mesh with connectivity, not just a bare point set. That distinction matters for anyone working on shape reconstruction, because the network learns to deform a set of surface patches (the 'atlas' parameterization) so the output is a watertight-ish surface you can texture and use downstream. The code is organized around two clear tasks: an autoencoder for point-cloud completion and a single-view reconstruction path, with a demo, training scripts, and pretrained models, so it is a practical starting point for reproducing the paper or adapting the architecture. Teams exploring 3D generation pipelines will find it a useful baseline to compare against newer methods, though it is research code built for Python 3.6 and older PyTorch, so expect to modernize dependencies before running it.