We previously developed a neural network which estimates the chromaticity of the illumination under which a given image was taken [1]. This provides colour constancy since, given the chromaticity estimate, the image pixel chromaticities can be converted via a diagonal transformation to what they would be under a canonical illuminant. In tests on synthetically generated and real scene images, the accuracy of the illumination-chromaticity estimate generally surpassed that of most existing colour constancy algorithms; however, the errors obtained with real images were significantly larger than those for the synthetic ones. After experiments with adding noise to the synthetic data, we concluded that there was a more fundamental problem than simply the influence of noise which remained to be explained. We hypothesized that specular reflection was causing the problem, so we modeled the specular reflection in the training set. The errors dropped by more than 20 percent.
Full text (gzipped postscript)