Abstract
The spectrum of the light reflected from a scene into the eye is the product of the scene's spectral reflecting properties and the spectrum of the illumination. As the illuminant changes, the excitations in each class of cone receptors of the eye change. A simple but accurate estimate of these changes in excitations is provided by the coefficient rule of von Kries, which conventionally incorporates two assumptions: that cone excitations depend on activity only within each cone class and that this dependence constitutes a simple scaling. Being able to predict the effects of an illuminant change allows it to be discounted as part of achieving an invariant perception of surface color, that is, color constancy. Both assumptions are important in modeling the mechanisms of color constancy. Although accounting for almost all of the variation in cone responses, von Kries' rule does show some systematic departures from proportionality. The aim of the present work was to test whether a non-parametric approach to predicting cone excitations, that is, one that does not depend on a particular parametric model of the effects of illumination, might be more accurate. Computer simulations were performed with hyperspectral images of natural scenes under separate illuminants drawn from combinations of sunlight, sky light, and filtered daylight transmitted through the forest canopy. Vegetated scenes were used rather than non-vegetated scenes as they were expected to reveal greater deviations from von Kries' rule. It was found that a non-parametric model based on locally weighted regression gave a significantly better fit than von Kries' scaling, suggesting that the departures from proportionality, although small, might be important. The improved performance of non-parametric fitting was achieved without compromising the basic assumption that excitations in each cone class depend on activity only within that class.
Supported by EPSRC Grant No. EP/C003470/1.