Reproduce “Intrinsic Single-Image HDR Reconstruction” 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
Sebastian Dille, Chris Careaga, Yağız Aksoy
Researchers from Simon Fraser University developed a method for single-image High Dynamic Range (HDR) reconstruction by remodeling the problem within the intrinsic image domain. This approach separates luminance range extension in the shading component from color recovery in the albedo component, achieving state-of-the-art accuracy on the SI-HDR benchmark with an HDR-VDP3 score of 8.96 JOD and producing high-fidelity HDR images.
Key insights
- Decomposing the LDR-to-HDR problem into intrinsic shading and albedo reconstruction simplifies the task for neural networks, allowing them to model distinct degradation patterns more effectively than monolithic approaches.
- Leveraging the physical separation of illumination (shading) and surface reflectance (albedo) allows for more stable and accurate learning: extending luminance range is handled in the shading domain, and color recovery in the albedo domain.
- The use of inverse representations for shading and HDR images, along with auxiliary losses that enforce consistency with the intrinsic image model, stabilizes network training and improves the fidelity of reconstructed details and colors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.