Re-Evaluating Colour Constancy Algorithms

We present a re-evaluation of previous experimental data for five different colour constancy algorithms, based on experiments on real and synthetic images. Our work is motivated by the observation that previous analysis of algorithm performance is flawed because it uses inappropriate statistical measures of performance. We discuss these flaws in detail and suggest more appropriate statistical tests. We show that using these tests conclusions as to the relative performance of algorithms are significantly changed as compared to the original analysis of the data. In particular we conclude that the performance of two algorithms: Gamut mapping and color by correlation is statistically equivalent and significantly better than the three other algorithms tested (Max-RGB and two versions of Grey-world).