Color Construction * Backgrounds and Illuminants: the Yin and Yang of Color Constancy

Over a wide range of viewing conditions, our experience of an object’s color varies little. This “color constancy” is a striking achievement of human vision, and raises a much-debated question: What is the proper theoretical framework for understanding color constancy? Several approaches have been considered. The most common has been to estimate the spectral composition of the illuminant, and to compensate for it while estimating the colors of object surfaces. An example of this approach is any model based on a variant of the “Grey World” hypothesis: the assumption that the average reflectance of surfaces in a visual scene is grey. This hypothesis allows one to estimate the spectral composition of the illuminant by simply computing how the space-averaged light from a scene deviates from grey. Another approach uses the assumption that illuminants and reflectances in nature can adequately be represented by linear models of low dimension, say two or three dimensions each. This assumption, when true, allows one to compute the reflectances and illuminants in a scene, thereby giving color constancy. In his paper “Backgrounds and illuminants: The yin and yang of color constancy” Richard Brown pursues a third approach, which urges careful attention to the evolutionary constraints on perception, and therefore to the ecological properties of the environment to which perception, and in particular color perception, might be adapted. Careful attention to ecology can uncover shortcomings of the standard approaches. Brown notes that careful study of the reflectances of natural scenes reveals that the Grey World hypothesis is, in general, false. Furthermore, careful study of reflectances and illuminants in nature reveals that they are not adequately represented by linear models of low dimension. Although the standard approaches are attractive for their computational simplicity, that simplicity apparently derives from unrealistically simplified representations of the natural ecology of reflectances and illuminants, and therefore such approaches are unlikely to perform adequately for natural scenes. They are, in consequence, unlikely to be adequate models for human visual performance.