A Hybrid Strategy For Illuminant Estimation
Targeting Hard Images

Roshanak Zakizadeh, Michael S. Brown and Graham D. Finlayson,
CPCV workshop at ICCV 2015
[paper (preprint)] [slides]

List of images from Gehler-Shi data-set captured by Canon5D camera which are classified by the hybrid method
as hard to be handled by simple algorithms such as White-Patch can be downloaded from here .

Images recommended to be removed from the Gehler-Shi data-set

Based on the analysis in our paper, we found that the majority of color constancy algorithms fail on the following nine images. On visual inspection, we found that these images contain two distinct ations (both indoor and outdoor). The ground truth illumination estimation for these scenes are from the color checker chart that is illuminated by a single illumination only. We recommend the removal of these images from the dataset for future research. The nine images are shown below, as well as the original Gehler-Shi datasets with the images removed.

Nine Images


Click here to download the updated set of raw images for Canon1D.

Click here to download the updated set of raw images for Canon5D.

Click here to download the zip file containing ground truth illuminants of both cameras provided in a matlab .mat file and the colour chart coordinates. You can also find a .mat file with the "id" and "name" of false images we removed from the dataset (please refere to the readme.txt file in the zip folder for guidance).