DehazeGS: Seeing Through Fog with 3D Gaussian Splatting

TL;DR

DehazeGS introduces a novel 3D Gaussian Splatting framework to dehaze foggy multi-view scenes. It learns a Gaussian depth-to-transmission mapping and jointly estimates atmospheric light, enabling explicit physical fog modeling within a 3D scene representation. A loss suite with DCP/BCP priors and pseudo-depth supervision guides accurate dehazing while preserving depth and 3D consistency. Across synthetic and real datasets, DehazeGS achieves state-of-the-art rendering quality with significantly faster performance than NeRF-based approaches, demonstrating strong practical impact for robust multi-view dehazing.

Abstract

Current novel view synthesis methods are typically designed for high-quality and clean input images. However, in foggy scenes, scattering and attenuation can significantly degrade the quality of rendering. Although NeRF-based dehazing approaches have been developed, their reliance on deep fully connected neural networks and per-ray sampling strategies leads to high computational costs. Furthermore, NeRF's implicit representation limits its ability to recover fine-grained details from hazy scenes. To overcome these limitations, we propose learning an explicit Gaussian representation to explain the formation mechanism of foggy images through a physically forward rendering process. Our method, DehazeGS, reconstructs and renders fog-free scenes using only multi-view foggy images as input. Specifically, based on the atmospheric scattering model, we simulate the formation of fog by establishing the transmission function directly onto Gaussian primitives via depth-to-transmission mapping. During training, we jointly learn the atmospheric light and scattering coefficients while optimizing the Gaussian representation of foggy scenes. At inference time, we remove the effects of scattering and attenuation in Gaussian distributions and directly render the scene to obtain dehazed views. Experiments on both real-world and synthetic foggy datasets demonstrate that DehazeGS achieves state-of-the-art performance. visualizations are available at https://dehazegs.github.io/

Figures (5)

• Figure 1: DehazeGS can generate accurate rendering results for scenes with participating media $(b)$. By learning disentangled representations for the participating media and the underlying scene, it simultaneously recovers the clear scene $(c)$ and obtains accurate depth estimation $(d)$.
• Figure 2: DehazeGS overview, we first obtain the Gaussian distributions in foggy scenes and perform alpha blending on the transmission of each Gaussian distribution to render the transmission map, which is guided and optimized using DCP and BCP priors. We utilize pseudo-depth maps as prior information for depth estimation when optimizing each input image.
• Figure 3: Qualitative comparison of novel view synthesis (input foggy images and render clear images under novel views unseen by the model) results on real datasets, the images of DehazeNeRF are taken from its original paper. We encourage readers to refer to the supplementary materials to view rendering results from more views.
• Figure 4: Qualitative comparison of dehazing results on the synthetic foggy dataset. Our method exhibits finer texture details and is closer to the ground truth (GT) compared to existing methods. In the supplementary materials, we provide other views and rendering results of the Waymo dataset.
• Figure 5: Qualitative ablation results show progressive improvements from vanilla GS to the complete model, with enhanced details and dehazing in distant and nearby scenes.