This website presents work about Color Constancy and holds the additional material to the CVPR 2008 paper Bayesian Color Constancy Revisited (conference poster). The main work for the paper was done during an internship of the first author at the Vision Group at Microsoft Research Cambridge. You can download the dataset and the code used for the experiments in the paper. In case of questions please write an email to the contact person.
Color constancy is the tendency to perceive surface color consistently, despite variations in ambient illumination. In this work we revisted the work presented in Bayesian Color Constancy with Non-Gaussian Models and investigated the use of more precice reflectance priors for Color Constancy. We collected a new dataset for Color Constancy consisting of different scenes, both indoor and outdoor taken under different illuminations. In each scene a Gretag MacBeth Color Checker Chart was placed such that it was illuminated by the main scene illuminant and thus its color could be retrieved. The code used for the experiments in the paper is also available.
Important: Different versions of this dataset exists. In case you are interested in a linear re-processing of the images I strongly recommend that you visit http://colour.cmp.uea.ac.uk/datasets/reprocessed-gehler.html and use the data made available there. The data you find there should be the one you want to work on! This website will be updated soon for a more complete description of the processing and different versions. Until then please visit the site mentioned above and use this site only if you are interested in the original RAW files.
What follows is the original description, this will be updated! Some example scenes of the dataset are shown below. The data is available in Canon RAW format free of any correction. For the experiment we use the Canon Digital Photo Professional Tool to convert the RAW images to TIFF format using the AutoWhitebalance setting of the camera. We also corrected for chromatic abberation in this step. The result was downsampled to 20% of the original size. You can download the RAW images split in the archives
This dataset has been re-processed by other researchers. To facilitate a fair comparison between differrent methods I recommend that you use the data provided here: