Reproduce “Invariant Descriptors for Intrinsic Reflectance Optimization” on OpenResearch

About this paper

Anil S. Baslamisli, Theo Gevers

Researchers from the University of Amsterdam introduce a physics-based approach for intrinsic image decomposition that leverages photometric invariant cross-color ratios to separate reflectance and shading. This method achieves improved performance on various datasets, notably reducing shadow artifacts and producing more robust reflectance maps compared to previous unsupervised and some supervised methods.

Key insights

• Physics-based invariant descriptors, such as cross-color ratios derived from fundamental reflection models, can substantially improve intrinsic image decomposition by distinguishing material changes from illumination effects.
• Unsupervised, learning-free approaches leveraging fundamental photometric principles can achieve performance competitive with or superior to supervised deep learning models in challenging scenarios like shadow removal.
• Robustness to strong shadows and varying illumination is enhanced by adaptively determining clusters and incorporating physical invariants, correctly identifying illumination-induced variations versus true reflectance changes.

Summary by alphaXiv.