SeaThru-NeRF adds the water to the NeRF rendering equation. Its rendering model follows the Sea-thru image formation model, and the architecture learns the scene and the medium’s parameters together. Because the medium is modelled separately, the trained field can render the scene with the water removed, recovering colour and depth of distant objects that the medium had largely hidden.
Findings
- The abstract shows photorealistic new views underwater on simulated and real scenes, plus clear renderings with the medium taken out. Code and datasets are released.
- Its real scenes became the standard benchmark that later splatting papers report on (li-2025-watersplatting, jiang-2025-rusplatting).
Limits
- Abstract only so far. NeRF training is slow and rendering is not real time; the splatting follow-ups attack both.
- Assumes camera poses from structure from motion.
Relevance
Later “model the water, then remove it” reconstruction work compares against it. Its benchmark scenes are fairly clear and shallow, a gap that alvareztunon-2026-gs-underwater tests directly. See underwater perception.
Links
- arXiv
- Subfield: Field robotics, Mapping and localisation