Print
Univis
Search
 
FAU-Logo
Techn. Fakultät Website deprecated and outdated. Click here for the new site. FAU-Logo

Color Constancy under Non-Uniform Illumination

Color constancy algorithms very often assume globally uniform illumination. However, in a number of scenes, this assumption is violated. Consider, for instance, an indoor scene with opened windows. Or a picture taken at night, with flash light. To achieve good results in such situations, a localized estimator for the illuminant color is required.

In our paper Initiates file downloadColor Constancy under Non-Uniform Illumination: Can Existing Algorithms Work?, we experiment with existing single-illuminant algorithms on multi-illuminant images. We segment the images, and apply off-the-shelf algorithms on these smaller regions. Preliminary results on a small dataset (Opens internal link in current windowavailable here) are encouraging. However - do we really want to 'recycle' single-illuminant algorithms? Or should we focus on specialized multi-illuminant algorithms?

The source code to this method is available, and Initiates file downloadcan be downloaded here. For installation instructions, see the file README. Note that the code comes with no warranty, and is licensed under GPL v3. If you use it in your own work, please reference our paper as

Michael Bleier, Christian Riess, Shida Beigpour Eva Eibenberger, Elli Angelopoulou, Tobias Tröger, Andre Kaupp: "Color Constancy under Non-Uniform Illumination: Can Existing Algorithms Work?", IEEE Color and Photometry in Computer Vision Workshop, Barcelona, Nov. 2011.

 If you have any questions or comments, please do not hesitate to contact Opens internal link in current windowChristian Riess.

Results

Upon reevaluating the code, our results improved a bit, in particular the fused estimators got a bit better. This was due to a slight inaccuracy in the handling of superpixels without any estimate. Below are the updated tables.

AlgorithmMeanMedianRMSMax.
Do nothing10.510.111.921.6
White patch Retinex5.95.17.221.6
Gray World5.04.95.614.4
1st order Gray Edge14.714.315.529.9
2nd order Gray Edge13.413.214.129.3
Gamut Mapping (max)6.05.27.421.3
Gamut Mapping (mean)6.66.57.315.0
Bayesian Color Constancy6.54.88.120.7
Gradient tree boosting3.83.24.616.1
Random forest regression3.73.24.514.2