← Back ICRA 2026

GS-UVCE: Gaussian Splatting-Driven Unsupervised Visual Consistency Enhancement for Underwater 3D Scene Reconstruction

Xiang Li, Chi Li, Yiming Xu, Yan Zhuang

PDF

AI summary

Key figure (auto-extracted from paper)

GS-UVCE achieves high-fidelity, multi-view consistent 3D reconstruction of degraded underwater scenes by jointly modeling water medium effects and illumination within an unsupervised Gaussian Splatting framework.

Underwater 3D reconstruction Gaussian Splatting Unsupervised enhancement Visual consistency Medium modeling Autonomous underwater vehicles

Problem

Underwater 3D reconstruction is hindered by optical distortions, scattering, and inconsistent lighting across views, while existing enhancement methods are typically supervised, single-view, and lack multi-view consistency.

Approach

The authors propose GS-UVCE, an end-to-end unsupervised framework that uses a Medium-MLP to model water attenuation and backscattering, and a Light-MLP to adaptively correct viewpoint-dependent illumination, guided by depth regularization for geometric consistency.

Key results

First end-to-end framework unifying unsupervised visual consistency enhancement with 3DGS
Novel Medium-MLP and Light-MLP modules for joint medium and illumination modeling
Superior reconstruction fidelity and visual consistency over SOTA methods on four public datasets
Effective mitigation of distance-dependent attenuation and scattering via depth regularization

Why it matters

Enables reliable, real-time 3D mapping and navigation for autonomous underwater vehicles operating in challenging, degraded environments.

Abstract

Underwater 3D scene reconstruction is critical for the operation of underwater robotics, yet remains highly challenging due to the semi-transparent water medium, which introduces optical distortions, light scattering, and severe visibility degradation. Therefore, effective underwater image enhancement is a prerequisite for reliable reconstruction. How- ever, existing approaches typically enhance individual views with pre-trained models before reconstruction, leading to poor generalization and inconsistent multi-view results. To address these limitations, we propose GS-UVCE, an end-to-end framework for Gaussian Splatting-driven Unsupervised Visual Consistency Enhancement. GS-UVCE incorporates a Medium- MLP to model water-medium effects and a Light-MLP to adap- tively correct illumination, ensuring illumination consistency. Furthermore, depth regularization is introduced to preserve geometric consistency under varying scene conditions. Extensive experiments on multiple underwater datasets show that GS- UVCE consistently outperforms SOTA methods, achieving supe- rior reconstruction fidelity and visual consistency enhancement. Code: https://github.com/CharlyX-Lee/GS-UVCE

Index terms

Visual Learning Deep Learning for Visual Perception RGB-D Perception