Abstract

3D Gaussian Splatting (3DGS) have set new benchmarks in reconstructing scenes from multi-view images. However, its application to scenes with scattering media like water or fog has been underexplored. Reconstructing images or scenes affected by scattering media is an ill-posed problem, as rendering the image requires knowing multiple physical attributes. In this paper, we propose a novel pipeline, UW-3DGS, to efficiently approximate the physical attributes of scattering media by learning an underwater image formation model and reconstructing underwater 3D scenes without water interference using a proposed Self-Pruning Supervision Branch with 3DGS. Experiments on real underwater scenes demonstrate its ability to render State-of-The-Art high-quality novel views of underwater scenes (with water optionally present), including essential details like seabed topography and vegetation, which are crucial for marine engineering. Quantitatively, our method is approximately 100 times more efficient in training and achieves a PSNR of 0.836 higher than SeaThru-NeRF, providing a more effective pipeline for underwater scene reconstruction and potential robotics applications.