Robust Low-light Scene Restoration via Illumination Transition

Ze Li, Feng Zhang, Xiatian Zhu, Meng Zhang, Yanghong Zhou, P. Y. Mok

Abstract

RoSe (Robust Low-light Scene Restoration) is a novel framework designed to synthesize high-quality, normal-light novel views from noisy, low-light multiview images. By reframing low-light enhancement as a 3D "illuminance transition" problem rather than a 2D image-processing task, RoSe achieves superior denoising and multiview consistency without requiring paired training data. This approach addresses the fundamental limitations of existing Neural Radiance Fields (NeRFs) in low-light conditions, such as geometric hallucinations and color inconsistency.

Introduction

Capturing high-quality 3D scenes in low-light environments is notoriously difficult. Standard NeRFs struggle because they rely on direct color optimization; when inputs are dominated by noise and shadows, models often misinterpret noise as geometry, leading to severe artifacts. Existing solutions generally fall into two categories: 2D pre-processing, which often causes "flickering" across different views, and end-to-end NeRF variants that frequently suffer from color distortion or the "fog" effect (misinterpreting dark air as solid geometry). RoSe aims to bridge this gap by creating a physically grounded way to transition from low-light observations to normal-light representations within a 3D volume.

Methodology

The core innovation of RoSe is its dual-branch architecture, which treats the relationship between low-light and normal-light as a specialized rendering task.

  • Dual-Branch Architecture: The model consists of a Viewer-Centered branch representing the "normal-light" scene and a World-Centered branch that estimates the Illuminance Transition Field. This field learns the transformation required to turn a low-light point into a normal-light point consistently across all camera positions.
  • Low-Rank Denoising (LRD) Module: RoSe exploits the physical property that natural illumination is typically "smooth" (low-rank), while sensor noise is "random and spiky" (high-rank). The LRD module filters out noise by projecting features into a low-rank subspace before final rendering.
  • Unsupervised Learning: The framework is trained without ground-truth bright images, utilizing a Mean Squared Error (MSE) loss on low-light inputs and an Illumination Correction (IC) loss to reach target brightness levels.
Overview of the RoSe architecture and the Low-Rank Denoising (LRD) module.

Key Insights

  • Transition over Decomposition: Unlike Retinex-based methods that attempt to decompose images into reflectance and illumination—an ill-posed problem—RoSe estimates the stable ratio (transition) between light levels.
  • Noise as High-Rank Interference: By treating denoising as a rank-reduction problem within the 3D field, the model avoids the blurring common in 2D denoising filters.
  • 3D Consistency: Because the illuminance transition is tied to 3D coordinates rather than 2D pixels, the resulting restored scene is perfectly consistent across all viewpoints, eliminating flickering artifacts.

Experimental Results

RoSe was evaluated on the LOM (Low-light Object Multiview) dataset, consisting of five distinct scenes captured under varying exposure settings. Quantitatively, RoSe significantly outperformed state-of-the-art models, achieving a mean PSNR of 26.28, compared to 23.13 for LLNeRF and 19.87 for Aleth-NeRF.

Qualitatively, visual results demonstrate that RoSe preserves sharp edges and realistic colors in areas where other models produce muddy textures or severe color shifts. Ablation studies confirmed that the LRD module is essential; removing it caused the PSNR to drop from approximately 26 to 21, validating the low-rank assumption for 3D denoising.

Qualitative comparison of RoSe against state-of-the-art low-light enhancement methods.

Conclusion

While RoSe currently assumes Lambertian surfaces and requires per-scene training, it represents a major advancement in robust 3D scene restoration. Future work will focus on integrating the framework into faster NeRF variants like Instant-NGP and testing on more diverse, large-scale outdoor environments.

Summary updated

BibTeX
@misc{li2025robust,
  author = {Ze Li and Feng Zhang and Xiatian Zhu and Meng Zhang and Yanghong Zhou and P. Y. Mok},
  title = {{Robust Low-light Scene Restoration via Illumination Transition}},
  year = {2025},
  eprint = {2507.03976},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.03976},
}
Figure