Abstract
Lightness and brightness have extensive research literatures, but their relationship is controversial. We used decision spaces to characterize them and to test computational models. In Experiment 1, we used a custom-built apparatus where adjustable reflectance patches were visible through two apertures, and illumination at the two apertures could be set independently. On each trial, reflectance and illuminance at the reference aperture were set to one of three settings. Reflectance and illuminance at the test aperture were randomly set to +/- 50% of the values at the reference aperture. In the lightness and brightness conditions, observers judged which aperture had a higher reflectance or luminance, respectively. For each of the three reference stimuli, we plotted the probability that the observer judged the test stimulus as lighter (or brighter), as a function of test reflectance and illuminance. Each such decision space was approximately divided in two by a straight line whose orientation varied across conditions. In the lightness task, the decision spaces were consistent with partial lightness constancy, with Thouless ratios around 0.80. In the brightness condition, Thouless ratios were lower, but decision spaces still indicated judgements closer to reflectance than to luminance judgements. In Experiment 2, we repeated this procedure with a rendering of the same apparatus on a monitor. Decision spaces were similar to those in Experiment 1, but indicated judgements more strongly influenced by luminance. Finally, we simulated computational models of lightness and brightness: ODOG, a high-pass model, a contrast normalization model, and two retinex models. All models’ decision spaces were highly inconsistent with those from human observers. We conclude that (a) lightness and brightness judgements are more similar than expected from previous work, (b) brightness is nothing like an estimate of luminance, and (c) current computational models can fail on even simple lightness and brightness judgements.