Reproduce “LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zefan Qu, Ke Xu, Gerhard Petrus Hancke, Rynson W.H. Lau
Researchers from City University of Hong Kong introduce LuSh-NeRF, an unsupervised framework that reconstructs high-quality Neural Radiance Fields from challenging hand-held low-light photographs afflicted by low visibility, sensor noise, and camera motion blur. The method sequentially decouples these degradation factors, leading to visually sharp, clean, and naturally lit novel view syntheses.
Key insights
- Noise in low-light images typically remains sharp despite camera motion, indicating an implicit causal sequence of degradation: low visibility, then noise, then blur.
- Decoupling view-inconsistent noise from the consistent 3D scene representation is crucial for robust NeRF reconstruction from degraded inputs, which is facilitated by multi-view consistency supervision.
- Frequency filtering combined with gradient detachment allows the blur estimation module to accurately predict camera trajectories by mitigating interference from high-frequency noise.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.