Uncertainty estimation in color constancy

Computational color constancy is an under-determined problem. As such, a key objective is to assign a level of uncertainty to the output illuminant estimations, which can significantly impact the reliability of the corrected images for downstream computer vision tasks. In this paper we present a formalization of uncertainty estimation in color constancy, and we define three forms of uncertainty that require at most one inference run to be estimated. The defined uncertainty estimators are applied to five different categories of color constancy algorithms. The experimental results on two standard datasets show a strong correlation between the estimated uncertainty and the illuminant estimation error. Furthermore, we show how color constancy algorithms can be cascaded leveraging the estimated uncertainty to provide more accurate illuminant estimates.