A Feasibility Study on Novel View Synthesis of Underwater Structures using Neural Radiance Fields

We explore Neural Radiance Fields (NeRFs) for synthesizing novel views of underwater structures. This learning-based approach relies on a sparse set of camera views to model the 3D geometry of underwater structures and scenes. Real-world underwater scenes exhibit significant temporal variations, introducing challenges in maintaining visual consistency. We investigate three NeRF implementations: 1) nerfacto, which rep-resents the baseline; 2) nerfacto with transient embed dings and 3) nerfacto with a robust loss, which are designed to deal with scene inconsistencies. We evaluate these implementations using datasets collected in 1) a controlled environment and 2) a real underwater setting. The modified implementations consistently outperform the original nerfacto across both datasets. The performance im-provements are particularly pronounced in the dataset obtained from the real underwater setting where scene inconsistencies are more prevalent. This underscores the importance of robustifying NeRF implementations to ensure consistent performance in the challenging underwater environments.