Abstract
The brightness of an image region depends on its local luminance as well as its surrounding context. For equiluminant patches embedded in different surrounds this can lead to brightness differences, such as in White's effect. Some of these context dependent lightness effects have been modeled computationally, but mainly for stereotyped stimuli where target luminance is at gray values intermediate to the black and white surround elements. How models respond to various target luminances, and how that affects predictions of the context effect, remains unclear. Here we study if the ODOG-family of multiscale spatial filtering models of brightness perception captures the relationship between target luminance and brightness and its context interaction for White’s effect. Stimuli consisted of square-wave carrier gratings with two targets placed on the black or white phases. Across stimuli, the luminances of these target patches were varied from black to white. Average model response in target regions was taken as the predicted brightness. The model transfer functions linking the target intensity and predicted brightness were straight lines for each target patch. An additive shift between the two transfer functions indicates that the model correctly predicted the direction of White’s effect, and that the effect’s magnitude did not vary with target luminance. However, humans do observe White’s effect varying with target luminances (Vincent, 2017). The underlying perceptual scales of target brightness as a function of target luminance — estimated using Maximum-Likelihood Conjoint Measurement (Aguilar, Vincent, Maertens, 2022) — are two nonlinear functions, which differ by both a shift and a change in shape. Passing the output of the ODOG models through a nonlinearity did produce nonlinear transfer functions, but still only produced an additive shift between the targets. Instead, introducing nonlinear steps in the model divisive normalization component may be necessary to better match the perceptual scales.