NeuralRecon

Why we included this project

NeuralRecon is a reference implementation for learning-based 3D reconstruction from a monocular video stream with known camera poses. Instead of estimating depth per frame and fusing later, it builds a coherent dense surface in real time using sparse TSDF volumes and a gated-recurrent fusion module. The repo includes training and inference code, data prep scripts, pretrained ScanNet weights, and a demo that runs on custom ARKit captures, so you can see reconstructions from your own phone footage quickly. The pretrained model also holds up on outdoor scenes and low-texture environments beyond its training domain, which makes it a practical baseline for AR, robotics, or scanning work. Note the torchsparse dependency and Linux assumption before diving in.