Underwater 3D Scene Reconstruction Using Neural Rendering: A Survey with Emphasis on 3D Gaussian Splatting Framework

Underwater 3D reconstruction is crucial for au- tonomous underwater vehicles (AUVs) and marine surveys but is more challenging than terrestrial scenarios due to attenuation, backscatter, and uneven lighting. While traditional photogram- metric pipelines fail to mitigate such effects, recent Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) are promising, however, still break down similarly when applied directly to underwater imagery, as both assume a transparent propagation medium. This survey provides a structured review of methods that adapt NeRF and 3DGS specifically for under- water 3D reconstruction. By formalising the underwater image formation model for specific architectural options, we introduce a novel taxonomy of four adaptation strategies based on phys- ical model integration, distractor and dynamic scene handling, depth and geometry-focused design, and sparse-view robustness. Furthermore, NeRF and 3DGS-based methods are evaluated on benchmark datasets and quantitatively compared using standard metrics, including PSNR, SSIM, and LPIPS. Therefore, we identify the open challenges covering forward scattering, unified benchmarking, dynamic scenes, sparse-view robustness, water- type generalisation, foundation model integration, real-time de- ployment, GS-SLAM, 3D foundation model priors, and multi- modal sensing fusion. To our best knowledge, this survey is among the first dedicated, taxonomy-driven, equation-level surveys of underwater 3DGS methods, including recent approaches such as OceanSplat and RUSplatting, which were absent from prior reviews.

Recommended citation: Shubham Parab, Shaojun Bian, Yun Wu, Safa Tharib. (2026). "Underwater 3D Scene Reconstruction Using Neural Rendering: A Survey with Emphasis on 3D Gaussian Splatting Framework." 12th International Conference on Virtual Reality 2026. IEEE.
Download Paper | Download Slides