Creating clear 3D digital maps of underwater environments is notoriously difficult. Water absorbs and scatters light, turning clear images into murky, distorted scenes where traditional 3D scanning algorithms fail. To solve this, we present Swimm3R, a unified system designed specifically for accurate underwater 3D reconstruction.
Swimm3R takes advantage of AI models trained on clear, above-water scenes to understand 3D shapes, and pairs them with a specialized physics module. This module calculates how water bends and absorbs light, simultaneously mapping the camera’s exact path and restoring clean 3D point clouds from murky video. To render these scenes in photorealistic detail, we developed Underwater Beta Splatting (UWBS), an enhanced 3D modeling technique that uses flexible "Beta" math shapes and accounts for light scattering to build smooth, accurate surfaces.
To test our framework, we built the new Barbados underwater video dataset, created from videos collected during the field trials we have conducted in the Caribbean Sea of the coast of Holetown, Barbados since January 2017. In heavy underwater haze, Swimm3R reliably reconstructs coherent seafloor landscapes. Compared to the current top method (WaterSplatting), Underwater Beta Splatting improves image rendering clarity by 1.47 dB PSNR and boosts camera tracking accuracy by 2.0% (rotation accuracy) and 2.4% (translation accuracy) under standard benchmark thresholds (RRA@15 / RTA@15).
Quantitative comparison of test-frame rendering quality on the Barbados dataset across feed-forward Structure-from-Motion (SfM) initializations and renderers.
Qualitative query-to-render feature matching for visual localization. UWBS produces more coherent matches across underwater appearance changes, showing that its rendered maps provide more reliable 2D correspondences and render-depth for downstream PnP pose estimation.
The paper can be found on arXiv and is currently under review.
A more detailed GitHub page can be found here with links to the code and dataset, and interactive results.
The BibTeX citation for the paper is currently this:
@article{kweon2026swimm3r,
title = {Swimm3R: Splatting with Medium-aware SfM for Underwater 3D Reconstruction},
author = {Kweon, Minseong and Sattar, Junaed},
journal = {arXiv preprint arXiv:2608.00950},
year = {2026}
}