Grounding Image Matching in 3D with MASt3R
Extends DUSt3R with a dense local-feature head, so one network gives both 3D pointmaps and accurate pixel matches between views — useful for calibration-free reconstruction and localization.在 DUSt3R 基础上加入稠密局部特征头，同一网络既输出三维点图又给出视角间精确的像素匹配，适用于免标定的三维重建与定位。
BibTeX
@misc{leroy2024grounding,
title = {{Grounding Image Matching in 3D with MASt3R}},
author = {Vincent Leroy and Yohann Cabon and Jérôme Revaud},
year = {2024},
eprint = {2406.09756},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2406.09756},
}