MASt3R

MASt3R, from Naver Labs Europe (arXiv June 2024, ECCV 2024), takes DUSt3R's pointmap-regression paradigm and augments it with a dense local-feature head trained with a matching loss, plus a fast reciprocal matching scheme with theoretical guarantees. The result is a standalone method for camera calibration, pose estimation, and 3D reconstruction that significantly outperforms prior art on challenging matching benchmarks like Map-free localization.

Where DUSt3R was robust but limited in matching accuracy, MASt3R preserves that robustness while adding precision, and its MASt3R-SfM variant delivers a fully integrated, unconstrained structure-from-motion pipeline. Weights are openly downloadable.