Illumination Estimation Challenge


All 1707 images of the Cube+ dataset that is described and available at can be used as the training data. These images were all taken with the same Canon EOS 550D camera. The PNG files there contain minimally processed raw images with only basic debayering performed.

As for the test data, it will consist of several hundred PNG images taken with the same Canon EOS 550D camera that was used to create the images for the training data. The encrypted archive with these images will be available on this page one week before the submission deadline, while the password for decrypting the archive will be available one day before the submission deadline.

BONUS training data

Interested parties could find it potentially useful to use artificially created data. What they need in that case is the Croatian Paper (CroP) dataset generator described in detail in the paper "CroP: Color Constancy Benchmark Dataset Generator" that is currently available at arXiv. The resources for the generator are available at


The encrypted test data is available HERE. It consists of 363 images with the lower right rectangle being masked out. Since the SpyderCube position in the images can vary, the mask area is now enlarged. In the training set, it was recommended that the lower right rectangle used for masking should start at row 1050 and column 2050, but in the test set, it now starts at row 900 and column 1900. The encrypted ground-truth is provided as well.

The password for the archive with test images is kEt2xPR?^%rqGKDj4LR8Z*NK.

The password for the ground-truth is *^$5CwRxr#b?nq#TnD2Wtw5F.