DehazeGS: 3D Gaussian Splatting for Multi-Image Haze Removal

Abstract

Neural Radiance Fields (NeRF) have advanced 3D reconstruction by learning implicit representations of scenes from multi-view images, yet their effectiveness is limited in environments with scattering medium. Existing methods that incorporate scattering models into NeRF frameworks face issues with slow training speeds and high memory demands. This paper presents DehazeGS, a novel haze removal and reconstruction method based on 3D Gaussian Splatting (3DGS). Our approach integrates the Koschmieder scattering model into the 3DGS framework, enabling effective separation of objects and scattering medium. This method leverages a point-based representation to achieve high-quality scene reconstruction while significantly reducing computational and memory overhead. Experimental results on both synthetic and real datasets demonstrate that our method outperforms existing approaches in terms of dehazing quality and reconstruction performance, effectively synthesizing clear images from foggy scenes. Our findings suggest that integrating scattering models with 3DGS offers a promising solution for applications in adverse weather conditions.