AtlantisGS: Underwater Sparse-View Scene Reconstruction via Gaussian Splatting

摘要

Underwater scenes present significant challenges for modern 3D scene reconstruction techniques due to absorption, in-scattering, and out-scattering effects, which alter light transport and degrade reconstruction quality, especially under sparse-view conditions. We present AtlantisGS, a novel underwater scene reconstruction method, which only requires sparse-view inputs. It incorporates a scattering decomposition method that separates medium and object contributions during rendering, and a sparse Gaussian proliferation strategy that adaptively densifies the scene representation to improve structural accuracy. These components jointly enhance both geometric reconstruction and medium modeling, enabling accurate and efficient scene recovery with limited observations. Extensive experiments on real-world underwater datasets demonstrate that AtlantisGS outperforms existing NeRF- and 3DGS-based methods across various metrics. AtlantisGS achieves higher reconstruction fidelity with significantly fewer input views and real-time rendering capability. These results establish AtlantisGS as an effective solution for sparse-view underwater 3D scene reconstruction.