How Multi-Illuminant Scenes Affect Automatic Colour Balancing

Many illumination-estimation methods are based on the assumption that the imaged scene is lit by a single course of illumination; however, this assumption is often violated in practice. We investigate the effect this has on a suite of illumination-estimation methods by manually sorting the Gehler et al. ColorChecker set of 568 images into the 310 of them that are approximately single-illuminant and the 258 that are clearly multiple-illuminant and comparing the performance of the various methods on the two sets. The Grayworld, Spatio-Spectral-Statistics and Thin-Plate-Spline methods are relatively unaffected, but the other methods are all affected to varying degrees.