LITA-GS: Illumination-Agnostic Novel View Synthesis via Reference-Free 3D Gaussian Splatting and Physical Priors

113 2. To enhance the scene structure, we develop lighting-114 agnostic structure rendering based on the spatial struc-115 ture prior extracted by our introduced illumination-116 invariant physical prior extraction pipeline. 117 3. Moreover, a lightweight progressive denoising module is 118 proposed based on noise rendering to surpass the noise. 119 4. Extensive experiments demonstrate that our LIT-3D sig-120 nificantly outperforms SOTA NeRF-based methods with 121 much faster speed. Compared to combining exposure 122 correction methods with 3DGS, our LIT-3D achieves 123 superior performance with improved multi-view consis-124 tency. 125 2. Related Works 126 3D Scene Representation for NVS in adverse illumina-127 tion conditions: NeRF [16] has gained popularity for its 128 ability to generate photorealistic 3D views from limited data 129 using deep neural networks. Subsequent works have ex-130 tended NeRF for 3D reconstruction of scenes with chal-131 lenging lighting conditions. NeRF-W [15] addresses vari-132 able lighting and transient occlusions in unstructured image 133 collections by incorporating image-dependent radiance ad-134 justments and identifying and managing transient elements 135 within scenes. RawNeRF [17] proposes training NeRF di-136 rectly on RAW data can effectively handle noise in dark 137 scenes. Given a set of commonly used sRGB images cap-138 tured in low-light scenes, LLNeRF [21] decomposes the 139 color of 3D points into illumination-related view-dependent 140 and view-independent components during NeRF optimiza-141 tion, facilitating the enhancement of novel view images.