Reproduce “RAW Image Reconstruction from RGB on Smartphones. NTIRE 2025 Challenge Report” 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
Marcos V. Conde, Radu Timofte, Radu Berdan, Beril Besbinar, Daisuke Iso, Pengzhou Ji, Xiong Dun, Zeying Fan and 25 more
The NTIRE 2025 RGB-to-RAW Challenge evaluated methods for reconstructing linear RAW sensor images directly from smartphone sRGB photographs without requiring camera metadata. The challenge demonstrated that high-fidelity metadata-free RAW reconstruction is feasible, with the top-performing DBNet achieving an overall PSNR of 27.66 dB and showing strong generalization across diverse smartphone models.
Key insights
- A significant generalization gap exists between performance on target (known) devices and out-of-family (unknown) devices.
- Simpler and more efficient models frequently demonstrated better generalization capabilities on unknown devices compared to larger, more complex architectures.
- The variety of successful methods, ranging from dual-path processing to dynamic convolutions, indicates multiple viable architectural paths for RAW reconstruction research.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.