Empirical evaluation of computational models of lightness perception

Area of Science:

• Visual perception
• Computational neuroscience
• Color science

Background:

• Surface lightness perception is influenced by surrounding context, leading to illusions.
• Computational models aim to replicate human visual system mechanisms for lightness.
• Assessing these models against empirical data is crucial for understanding visual processing.

Purpose of the Study:

• To evaluate the performance of eight computational models in predicting human lightness perception.
• To compare model predictions with human participant data across diverse visual stimuli.
• To identify models that best capture the complexities of lightness illusions and contrast effects.

Main Methods:

• Eight computational models were applied to 13 distinct visual displays (11 illusions, 2 Mondrians).
• Model outputs were quantitatively compared with psychophysical data from 85 human participants.
• Model accuracy was assessed for various lightness phenomena, including simultaneous lightness contrast and specific illusions.

Main Results:

• HighPass and MIR models accurately predicted simultaneous lightness contrast (SLC) effects.
• ODOG variants and RETINEX models showed success with specific illusions like White's and Dungeon illusions.
• The FL-ODOG model demonstrated the highest overall accuracy, predicting results for most displays, but struggled with Reversed contrast illusion.
• Most models underperformed on Mondrian displays, which represent more naturalistic visual scenes.

Conclusions:

• No single model perfectly replicates human lightness perception across all tested conditions.
• FL-ODOG shows the most promise for simulating human lightness judgments, particularly for illusory displays.
• Further model development is needed to accurately account for performance on naturalistic stimuli like Mondrian displays.