UMB-GSfM: Underwater Marker-Based Global Structure-from-Motion

Traditional Structure-from-Motion (SfM) methods often struggle with robustness and accuracy in underwater environments due to the effects of refractive imaging. To address the challenge in underwater visual 3D reconstruction, we propose a novel marker-based global SfM framework tailored for underwater scenarios. Specifically, relative poses are first computed at a real scale between different image frames through inter-frame marker matching. Then a novel refractive imaging model-adapted reprojection error is introduced, which formulates the SfM problem as a bundle adjustment optimization. Additionally, a distance-based pruning method is developed to improve the accuracy of sparse 3D point cloud generation. By mitigating the risks of erroneous data association, the proposed framework effectively solves both the relative pose estimation and bundle adjustment problems using reliable nonlinear optimization solvers. Experimental results on several practical datasets demonstrate the superior performance of our method compared to state-of-the-art approaches.