Reproduce “NeRF in the Dark: High Dynamic Range View Synthesis from Noisy Raw Images” 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
Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron
RawNeRF enables high dynamic range (HDR) novel view synthesis directly from noisy, low-light raw images by reconstructing scenes in a linear HDR space. This approach allows for robust multi-view denoising and provides flexible post-rendering control over exposure and focus, extending NeRF's applicability to challenging real-world photographic conditions.
Key insights
- Optimizing NeRF directly in raw, linear HDR space allows it to implicitly perform highly effective multi-image denoising by averaging out zero-mean noise across numerous diverse views.
- Preserving the full dynamic range of a scene during reconstruction enables flexible post-rendering creative controls, such as arbitrary exposure adjustment, diverse tonemapping, and physically accurate synthetic defocus with realistic bokeh.
- Working with unprocessed raw sensor data avoids the detrimental effects of conventional camera image signal processing pipelines, which can distort noise distributions and lose detail in highlights and shadows.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.