The relationship of lightness illusions uncovered by individual differences and its advantage in model evaluation

Computational models that explain lightness/brightness illusions have been proposed. These models have been assessed using a simplistic criterion: the number of illusions each model can correctly predict from the test set. This simple method of evaluation assumes that each illusion in the set is independent; however, since lightness illusions are not independent of each other, an unbalanced test set may distort the evaluation of models. Moreover, evaluating models with a single value obscures where the model’s strength and weakness lies. We collected illusion magnitudes of various lightness illusions through two online experiments and investigated their correlations to identify underlying factors in these illusions. Experiment 1 identified three common factors reflecting assimilation, contrast, and White’s effect. Experiment 2, with a different illusion set, identified two factors reflecting assimilation and contrast, and detected high independence of an illusion caused by remote luminance gradients. We then examined three well-known models using the outcomes of the experiments. This model test showed that the redundancy in the illusion sets does not markedly skew the evaluation of the models, but the predictions by some models have a strong bias towards the contrast effect. This study clarified that correlations of illusion magnitudes provide valuable insights into both illusions and models, and highlighted the need to assess models based on processes underlying lightness perception, rather than focusing on individual illusions.