DUSt3R

DUSt3R (Dense and Unconstrained Stereo 3D Reconstruction), presented at CVPR 2024 with an arXiv preprint in December 2023, replaced the classical SfM plus multi-view-stereo pipeline with a single feed-forward network. It regresses 3D pointmaps directly from arbitrary image pairs with no camera calibration, no pose priors, and no correspondence search.

Built on standard transformer encoders and decoders leveraging pretrained models, it unified monocular and multi-view depth estimation and relative pose estimation, setting new records on both. Its global alignment strategy extends pairwise reconstructions to full image collections, and it became the foundation for an entire family of follow-ups including MASt3R, CUT3R, and VGGT. Code and weights are openly available.