AquaFlow is a monocular, streaming Gaussian splatting SLAM system for underwater video. A 3D vision foundation model fine-tuned on large-scale underwater data gives poses and pointmaps; a medium-guided strategy decides where to add new Gaussians as the map grows; and the scene is a hybrid of structured, distance-conditioned neural Gaussians plus a physics-inspired optical model for the water.
Findings
- Evaluated on 62 underwater trajectories from public benchmarks and web videos (abstract).
- Compared with WaterSplat-SLAM: 13.2% lower average localisation error and 4.74 dB higher PSNR.
Limits
- Abstract only so far. Web videos have no ground-truth trajectory, so I would want to see how localisation error was measured on those.
Relevance
Streaming is what an onboard map needs, and fine-tuning the pose model on underwater data addresses the same domain gap as Wat3R. See underwater perception and murai-2025-mast3r-slam.
Links
- arXiv
- Subfield: Field robotics, Mapping and localisation