TableDB dataset: A newly added underwater dataset includes scene captures from different distances and viewpoints.
Most useful features in underwater images are occluded by water. Consequently, underwater scenes present additional difficulties beyond the usual challenges of 3D reconstruction and rendering for in-air scenes. Naive applications of Neural Radiance Field methods (NeRFs) or Gaussian Splatting, while highly robust for in-air scenes, fail to perform underwater. Here we introduce Gaussian Splashing, a new method based on 3D Gaussian Splatting (3DGS), into which we incorporate an image formation model for scattering. Concretely, we introduce three additional learnable parameters to the rendering procedure, modify the depth estimation step for underwater conditions, alter the original loss term used in 3DGS, and introduce an additional loss term for backscatter. Our approach achieves state-of-the-art performance for underwater scenes and is highly efficient, with 3D reconstruction taking only a few minutes and rendering at 140 FPS, despite the complexities of underwater adaptation.