manhattan_sdf

Why we included this project

Researchers working on neural implicit surface reconstruction will find this a clean reference implementation of a CVPR 2022 method. It reconstructs dense 3D indoor scenes from multi-view images by pairing a neural signed distance field with the Manhattan-world assumption, which helps it recover clean geometry on low-textured planar surfaces where plain NeRF-style approaches tend to struggle. The code is compact and readable: training, mesh extraction, and evaluation all run from simple config files, and the repo ships processed ScanNet data plus baseline results so you can reproduce the paper's comparisons without rebuilding the data pipeline. Treat it as a research baseline and a starting point for extending indoor scene reconstruction rather than a production-ready tool, and check the custom non-commercial license before any commercial use.