UFSt3R: advancing uncalibrated multi-view 3D reconstruction via attention-guided masking and global geometric feedback

Abstract With the rapid advancement of computer vision, multi-view 3D reconstruction has become a cornerstone in real-world applications such as augmented reality and autonomous driving. However, existing approaches still suffer from structural redundancy, weak cross-view feature robustness, and limited modeling capability in textureless or occluded regions. To address these challenges, this paper proposes UFSt3R (uncalibrated fast stereo 3D reconstruction), a lightweight and calibration-free framework designed for robust and efficient 3D reconstruction. UFSt3R integrates a shared vision transformer encoder for unified multi-view feature extraction, a shared Transformer decoder with attention-guided masking to enhance cross-view matching selectivity and robustness, and an Inj3D global geometric feedback module to propagate multi-level structural cues, improving reconstruction consistency in challenging regions. Extensive experiments on multiple benchmark datasets demonstrate that UFSt3R consistently outperforms existing methods in unsupervised camera pose estimation, relative pose recovery, and dense 3D reconstruction, achieving superior accuracy and computational efficiency. Ablation studies further confirm that each proposed module contributes critically to the performance, establishing UFSt3R as a promising paradigm for robust, real-time, and calibration-free 3D reconstruction.