Kangxu Wang* · Shaofeng Zou* · Chenxing Jiang* · Yixiang Dai · Siang Chen · Shaojie Shen · Guijin Wang†
Underwater monocular SLAM is a highly challenging problem with applications ranging from autonomous underwater vehicles to marine archaeology. However, existing underwater SLAM methods struggle to generate high-fidelity rendered maps. We propose WaterSplat-SLAM, the first novel monocular underwater SLAM system to achieve robust pose estimation and photorealistic dense map construction to our knowledge.
Specifically, we combine semantic medium filtering with a dual-view 3D reconstruction prior to achieve underwater adaptive camera tracking and depth estimation. Furthermore, we propose a semantically guided rendering and adaptive map management strategy, combined with an online medium-aware Gaussian map, to model the underwater environment in a photorealistic and compact manner. Experiments on multiple underwater datasets demonstrate that WaterSplat-SLAM achieves robust camera tracking and high-fidelity rendering in underwater environments.