Abstract
Numerous computational models have been proposed to account for brightness/lightness illusions. To compare these models, researchers often start with a set of test illusions and then count the number of illusions in the set that each model can correctly predict. Models are considered successful if they can account for more illusions than other models; however, since lightness illusions are not independent of each other, one class of model may seem better than another class of model, but the ranking of may be due to a stimulus set that overrepresents correlated illusions. Here, we collected the magnitudes of various lightness illusions through two online experiments and then examined the response with an exploratory factor analysis. We found that the illusions in typical test sets can be divided into three classes: assimilation illusions, contrast illusions, and Whites effect combination illusions. We also had one example of a contrast enhancement illusion (Agostini’s glare illusion) that seemed to be separate from the other three classes. We then examined three well-known computational models (ODOG, LODOG, and FLODOG) using the obtained information about relationships of the illusions. We show that the assumption of illusions’ independence does not markedly distort the relative evaluation of the models, but performances of some models are substantially unbalanced across the three classes. ODOG and LODOG tend to be better at contrast effects, while FLODOG has better balance across three classes, but all three models were not very good at accounting for the contrast enhancement example. The results are consistent with the idea that there are separate processes for assimilation and contrast and a higher order stage that combines these separate processes. The study highlights the need to assess model performance based on their explanation of underlying processes rather than focusing solely on individual illusions.