Can DNN models simulate appearance variations of #TheDress?

Abstract

The color appearance of #TheDress image varies across individuals. The color of pixels in the image distributes mostly in blue-achromatic-yellow color direction, and so are the perceived color variations. One of the potential causes is differences in the degree of perceiving light-blue pixels as a part of white clothing under a skylight, referred to as "blue bias." A deep neural network (DNN) application was used to simulate individual differences in blue bias, by varying the percentage of such scenes in the training-image set. A style-transfer DNN was used to simulate a "color naming" procedure by learning pairs of natural images and their color-name labels as pixel-by-pixel maps. The models trained with different ratio of blue-bias scenes were tested using the #TheDress image. The averaged results across trials showed a progressive change from blue/black to white (gray)/gold, indicating that exposure or attention to blue-bias scenes could have caused the individual differences in the color perception of #TheDress image. In an additional experiment, we manipulated the relative number of artificially blue- or yellow-tinted images, instead of varying the ratio of blue-bias scenes, to train the DNN. If the blue-bias scenes are equivalent with blue-tinted images of scenes taken under daylight, this manipulation should yield similar result. However, the resulting outputs did not produce a white/gold image at all. This suggests that exposure to skylight scenes alone is insufficient; the scenes must contain unequivocally white objects (such as snow, white clothing, or white road signs) in order to establish a "blue bias" in human observers.