simplerecon

Why we included this project

SimpleRecon is the reference PyTorch implementation of an ECCV 2022 method that turns a set of posed photographs into per-image depth maps for indoor 3D reconstruction. Where earlier multi-view stereo approaches relied on expensive 3D convolutions, it substitutes cheaper 2D ones, which trims memory and compute while keeping accuracy strong on ScanNet. If you work on depth estimation or scene reconstruction from posed images, the repo ships a complete training and evaluation pipeline, pre-trained weights, and a documented environment setup. Since the code is organized around standard MVS building blocks like cost volumes and feature extraction, it's also a useful place to start if you want to understand how modern multi-view stereo pipelines work rather than just run them. The license is limited to non-commercial and academic use, so check the terms before folding it into commercial work.