Can DNN models simulate appearance variations of #TheDress?

Area of Science:

• Vision Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• The color appearance of #TheDress image is highly variable among individuals.
• This variation is partly attributed to differences in perceiving blue pixels as white under skylight, termed "blue bias."

Purpose of the Study:

• To simulate individual differences in color perception using a deep neural network (DNN).
• To investigate the role of "blue bias" in the perception of #TheDress image.
• To determine if skylight exposure alone or the presence of white objects is crucial for establishing "blue bias."

Main Methods:

• A style-transfer DNN was trained on image datasets with varying percentages of blue-biased scenes.
• The DNN simulated color naming by learning image-label pairs.
• Models were tested on #TheDress image, and a separate experiment used artificially tinted images.

Main Results:

• DNNs trained with higher ratios of blue-biased scenes showed a shift in #TheDress perception from blue/black towards white/gold.
• Artificially blue-tinted images did not produce a white/gold perception, unlike blue-biased scenes.
• This indicates that unequivocally white objects in scenes are necessary to establish "blue bias."

Conclusions:

• Exposure to scenes with white objects under skylight is critical for developing "blue bias."
• Skylight alone is insufficient to induce the perceptual shift seen in #TheDress.
• DNN simulations can effectively model individual differences in visual perception related to environmental lighting cues.