WaterClear-GS: Optical-Aware Gaussian Splatting for Underwater Reconstruction and Restoration

TL;DR

WaterClear-GS tackles underwater 3D reconstruction and appearance restoration by embedding wavelength-dependent attenuation and scattering directly into 3D Gaussian primitives. It introduces a dual-branch rendering approach that jointly optimizes underwater and water-free views, augmented by depth-guided geometry regularization, perception-driven image loss, exposure control, spatially-adaptive regularization, and a physically guided spectral prior. The method achieves state-of-the-art results for novel view synthesis and underwater image restoration while maintaining real-time rendering speeds, and it is validated on public SeaThru datasets plus a new ShipWreck dataset. Crucially, it eliminates the need for auxiliary medium networks, offering a compact, interpretable, and efficient framework for underwater scene modeling. The work advances practical underwater 3D vision by delivering high-fidelity geometry and faithful color restoration under diverse water conditions.

Abstract

Underwater 3D reconstruction and appearance restoration are hindered by the complex optical properties of water, such as wavelength-dependent attenuation and scattering. Existing Neural Radiance Fields (NeRF)-based methods struggle with slow rendering speeds and suboptimal color restoration, while 3D Gaussian Splatting (3DGS) inherently lacks the capability to model complex volumetric scattering effects. To address these issues, we introduce WaterClear-GS, the first pure 3DGS-based framework that explicitly integrates underwater optical properties of local attenuation and scattering into Gaussian primitives, eliminating the need for an auxiliary medium network. Our method employs a dual-branch optimization strategy to ensure underwater photometric consistency while naturally recovering water-free appearances. This strategy is enhanced by depth-guided geometry regularization and perception-driven image loss, together with exposure constraints, spatially-adaptive regularization, and physically guided spectral regularization, which collectively enforce local 3D coherence and maintain natural visual perception. Experiments on standard benchmarks and our newly collected dataset demonstrate that WaterClear-GS achieves outstanding performance on both novel view synthesis (NVS) and underwater image restoration (UIR) tasks, while maintaining real-time rendering. The code will be available at https://buaaxrzhang.github.io/WaterClear-GS/.

Figures (10)

• Figure 1: Illustration of underwater imaging and our results. Top: Light from objects is affected by significant attenuation and scattering, which intensify with distance. Bottom: Our method enables high-quality rendering at over 160 FPS, effectively removes scattering effects, and restores true colors.
• Figure 2: Overview of the WaterClear-GS framework. Our method extends each Gaussian with water optical parameters $\beta^{D}$, $\beta^{B}$ and $B$. The dual-branch design simultaneously renders underwater images by applying these parameters and clear images by zeroing them out. Depth-guided enhancement guides the geometry optimization, while exposure constraint balances the dynamic range of restored color and spatially-adaptive regularization, together with spectral regularization, ensures physical plausibility of medium properties. Our framework ensures high-quality reconstruction and realistic color restoration.
• Figure 3: Qualitative results of the NVS task. We present rendered underwater images and their depth maps. The pseudo-depth maps are used as a reference. Zoomed-in regions (highlighted with red bounding boxes) illustrate detailed differences. Our method consistently preserves geometric structures and fine-level details in most cases.
• Figure 4: Qualitative results of the UIR task. Details are zoomed in and highlighted with red and yellow bounding boxes. In contrast to other methods, which often produce grayish, underexposed, or severely color-shifted results, our method restores images with more natural visual quality.
• Figure 5: Visualization of learned optical parameters with/without $L_p$ on the IUI3-RedSea scene.