Experimental Comparsion Between NeRFs and 3D Gaussian Splatting for Underwater 3D Reconstruction

In the last two decades, research on the field of visual-based 3D reconstruction has undergone a significant evolution especially in the area of deep learning. Typically Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS) are used for representations in high-fidelity view synthesis and have made rapid development. However, such techniques have not been fully evaluated in challenging conditions, such as the underwater domain while it is vital to human ocean exploration, archaeology, oil and gas industry and mine countermeasure missions. In this paper, we compare the performance of NeRFs and 3DGS techniques in the underwater reconstruction on the dataset QDLake-UW, collected at QianDao Lake Underwater by ourselves on different conditions.