Factorized Gaussian Splatting for View-Inconsistent Degradations

Xiangyu Wen, Guangchi Fang, Changhao Chen, Chris Xiaoxuan Lu, Bing WANG

Keywords: 3D reconstruction, Gaussian Splatting, Thermal, Underwater

Abstract: Multi-view reconstruction often assumes cross-view appearance consistency, which is violated in the presence of participating medium and sensor artifacts. These view-inconsistent degradations induce distorted geometry and unreliable appearance in scene modeling. In this study, we propose Factorized Gaussian Splatting (F-GS), a unified scene modeling framework that explicitly decomposes the scene into three complementary components: Geometry, Medium, and Residual. The geometry component is stabilized by a lightweight view-aware surface gain and an optional screen-space normal-consistency prior to consolidate scene structure. The medium component is represented by sparse yet large Gaussians to model light attenuation through participating medium. The residual component captures view-dependent sensor noise via residual Gaussians, compensating for sensor-induced variations. This factorization prevents view-inconsistent degradations from contaminating geometry and appearance, enabling more accurate scene modeling. We instantiate and evaluate our F-GS in three representative regimes, thermal long-wave infrared, underwater, and foggy scene, where the view-inconsistencies naturally occur. Our F-GS significantly improves novel-view synthesis quality and geometric accuracy over other baselines.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 11576