AbstractThe problem that color constancy models address is one of recovering the reflectances of a surface under different illuminants and spatial configurations. As models of the human visual system, they should recover surface reflectances when the human visual system recovers those reflectances and fail when the human visual system fails to recover those reflectances. Unfortunately, it is difficult to determine just when these models will recover surface reflectances and when they will fail. We introduce a theory of influence that quantifies the degree to which the lightness calculated for a given surface element is influenced by the receptor responses to other elements. Influence theory is shown to be a very useful tool for highlighting testable predictions that can discriminate among color constancy models without having to determine directly when they will recover surface reflectances. This technique for generating empirical predictions is demonstrated by studying the effects of choosing different normalizing functions on an asymptotic retinex algorithm.
