AUTHOR=Fan Xinnan , Wang Xiaotian , Ni Haonan , Xin Yuanxue , Shi Pengfei 

TITLE=Water-Adapted 3D Gaussian Splatting for precise underwater scene reconstruction

JOURNAL=Frontiers in Marine Science

VOLUME=Volume 12 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2025.1573612

DOI=10.3389/fmars.2025.1573612

ISSN=2296-7745

ABSTRACT=Underwater 3D reconstruction is essential for marine surveying, ecological protection, and underwater engineering. Traditional methods, designed for air environments, fail to account for underwater optical properties, leading to poor detail retention, color reproduction, and visual consistency. In recent years, 3D Gaussian Splatting (3DGS) has emerged as an efficient alternative, offering improvements in both speed and quality. However, existing 3DGS methods struggle to adaptively adjust point distribution based on scene complexity, often resulting in inadequate detail reconstruction in complex areas and inefficient resource usage in simpler ones. Additionally, depth variations in underwater scenes affect image clarity, and current methods lack adaptive depth-based rendering, leading to inconsistent clarity between near and distant objects. Existing loss functions, primarily designed for air environments, fail to address underwater challenges such as color distortion and structural differences. To address these challenges, we propose an improved underwater 3D Gaussian Splatting method combining complexity-adaptive point distribution, depth-adaptive multi-scale radius rendering, and a tailored loss function for underwater environments. Our method enhances reconstruction accuracy and visual consistency. Experimental results on static and dynamic underwater datasets show significant improvements in detail retention, rendering accuracy, and stability compared to traditional methods, making it suitable for practical underwater 3D reconstruction applications.