Reproduce “A Dataset of Multi-Illumination Images in the Wild” 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
Lukas Murmann, Michael Gharbi, Miika Aittala, Fredo Durand
Researchers from MIT CSAIL and Adobe Research present a large-scale, high-quality dataset of over 1000 indoor scenes captured under 25 distinct HDR lighting conditions, utilizing a portable bounce-flash system. This dataset enables deep learning models to achieve improved performance in single-image illumination estimation, relighting, and mixed-illumination white balance tasks.
Key insights
- Learning-based inverse rendering problems require explicit supervisory signals from multi-illumination data, which are missing in conventional single-illumination datasets.
- A portable, automated bounce-flash system can effectively capture large-scale, high-quality multi-illumination datasets of real-world indoor scenes, overcoming the limitations of traditional lab-based setups.
- Training deep learning models on this diverse multi-illumination dataset leads to substantially improved performance in challenging tasks like single-image illumination estimation, relighting, and mixed-illuminant white balance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.