A Curious Problem with Using the Colour Checker Dataset for Illuminant Estimation

摘要

In illuminant estimation, we attempt to estimate the RGB ofrnthe light. We then use this estimate on an image to correct for thernlight’s colour bias. Illuminant estimation is an essential componentrnof all camera reproduction pipelines. How well an illuminantrnestimation algorithm works is determined by how well it predictsrnthe ground truth illuminant colour. Typically, the ground truthrnis the RGB of a white surface placed in a scene. Over a largernset of images an estimation error is calculated and different algorithmsrnare then ranked according to their average estimationrnperformance. Perhaps the most widely used publically availablerndataset used in illuminant estimation is Gehler’s Colour Checkerrnset that was reprocessed by Shi and Funt. This image set comprisesrn568 images of typical everyday scenes.rnCuriously, we have found three different ground truths forrnthe Shi-Funt Colour Checker image set. In this paper, we investigaternwhether adopting one ground truth over another results inrndifferent rankings of illuminant estimation algorithms. We findrnthat, depending on the ground truth used, the ranking of differentrnalgorithms can change, and sometimes dramatically. Indeed, it isrnentirely possible that much of the recent ’advances’ made in illuminantrnestimation were achieved because authors have switchedrnto using a ground truth where better estimation performance isrnpossible.