OS1-1 How Multi-Illuminant Scenes Affect Automatic Colour Balancing

The usual first step in automatic colour balancing of digital imagery is to estimate the chromaticity of the illumination. Although there are some recent exceptions (Beigpour et al. 2014. Gijsenij et al. 2013), most illumination-estimation methods assume that the relative spectral power distribution of the illumination is constant throughout the scene. However, many scenes contain multiple illuminants with differing SPDs, and we investigate the effect this has on automatic colour balancing. Somewhat surprisingly, the Gehler-Shi data set of 568 images, which is widely used in evaluating competing illumination-estimation methods, contains many images of multiple-illuminant scenes: for example, indoor scenes that also include a window through which daylight is visible. Each image in the dataset contains an Xrite/Macbeth ColorChecker that is used to provide a ground-truth measure of the illumination’s ‘colour’. However, since many of the scenes do contain multiple illuminants, a single such measurement cannot possibly represent the colour of all the illuminants correctly, but rather must represent some sort of compromise. We investigate how much of an effect this has on a representative set of illumination-estimation methods; namely, MaxRGB, Greyworld, Shades-of-Grey, Edge-based, N-jet and Thin-plate-Spline. We manually sorted the original 568 images into two groups accordin