Purchase this article with an account.
Erica Dixon, Arthur Shapiro; Can high-pass filters remove the effects of illumination? A filter-based approach to understanding #theDress and other demonstrations of color constancy. Journal of Vision 2016;16(12):221. doi: 10.1167/16.12.221.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Because material reflectance remains invariant across illumination conditions, many theories of color constancy assume that an observer's perceptual goal is to infer material reflectance. According to these theories, one way to estimate reflectance is to discount the illumination in the scene based on the observer's previous experience. Here we explore an alternate approach based on the idea that global illumination usually affects low spatial frequency content and therefore can be removed from an image with a simple high-pass filter. We demonstrate that an object that appears to vary in color when photographed under blue, white, or yellow illumination does not change color when the images are filtered so as to remove the low spatial frequency information. The implication is that, to a first approximation, high spatial frequency content remains invariant to changes in global illumination and therefore can be used as a basis of color constancy. We propose that the cut-off frequency for the filter adjusts to the sizes of objects in the scene, which prevents desaturation. This filter approach can account for the Rubik's cube illusion (a canonical demonstration for color constancy). Lastly, we examine how high-pass filtering affects observer perceptions of #theDress. If high-pass filtering acts as a method for discounting illumination, then observers should be more likely to perceive the filtered versions of the dress as "blue-black." However, observers are more likely to report that the dress appears "white-gold" as more low spatial frequency content is removed. The results could be taken as support for the theory that observers are not trying to estimate material reflectance (a distal property) but are instead giving a response based on an image's high spatial frequency content (a proximal property). The results are consistent with other reports that suggest separate high and low spatial frequency processes for color constancy (Werner, 2014).
Meeting abstract presented at VSS 2016
This PDF is available to Subscribers Only