Color Constancy and Non-Uniform Illumination: Can Existing Algorithms Work?
📜 Abstract
The color and distribution of illuminants can significantly alter the appearance of a scene. The goal of color constancy (CC) is to remove the color bias introduced by the illuminants. Most existing CC algorithms assume a uniformly illuminated scene. However, more often than not, this assumption is an insufficient approximation of real-world illumination conditions (multiple light sources, shadows, interreflections, etc.). Thus, illumination should be locally determined, taking under consideration that multiple illuminants may be present. In this paper we investigate the suitability of adapting 5 state-of-the-art color constancy methods so that they can be used for local illuminant estimation. Given an arbitrary image, we segment it into superpixels of approximately similar color. Each of the methods is applied independently on every superpixel. For improved accuracy, these independent estimates are combined into a single illuminant-color value per superpixel. We evaluated different fusion methodologies. Our experiments indicate that the best performance is obtained by fusion strategies that combine the outputs of the estimators using regression.
✨ Summary
Summary
The paper investigates whether single-illuminant color-constancy algorithms can be adapted to scenes containing spatially varying illumination. It segments images into color-based superpixels, applies Gray World/Gray Edge, Gamut Mapping, and Bayesian color-constancy estimators independently within each region, and fuses their local predictions. The authors introduce a laboratory dataset containing four scenes recorded under 36 multi-illuminant conditions, using gray-painted versions of the scenes to obtain localized illumination ground truth.
The central result is that individual estimators degrade substantially on small, color-homogeneous regions, but regression-based fusion—especially random-forest regression—recovers much of the lost accuracy. On the non-uniform-illumination test set, random-forest fusion achieved a median angular error of 4.1°, compared with 4.6° for the best individual estimator reported in the paper. The approach therefore provides evidence that ensembles of locally applied conventional estimators can produce useful dense illuminant estimates without additional image features, although the authors identify computational cost, limited training data, and ground-truth quality as important limitations.
The paper’s influence is primarily methodological and foundational within later multi-illuminant color-constancy research. The authors’ institutional project page reports accompanying source code and data, and later work on conditional-random-field estimation, mixed-illumination datasets, and neural or generative multi-illuminant methods explicitly cites this paper as prior work. (www5.informatik.uni-erlangen.de) Subsequent research moved beyond the paper’s superpixel-plus-estimator-ensemble design toward specialized spatial models, learned CNN estimators, and larger multi-illuminant datasets; the paper is consequently best characterized as an early empirical demonstration that localized illumination estimation is feasible by adapting existing algorithms rather than as a final practical solution. (openaccess.thecvf.com)