Abstract
Our visual system is capable of achieving constant color perception across large changes in illumination. Consequently, a physically constant, colorimetrically neutral patch of light, appearing gray under a neutral illumination, will appear vividly colored if the observer has the impression that the illumination of the scene changed. We wanted to test whether we can decode the brain's responses to color with respect to appearance, rather than mere retinal cone stimulation. We presented observers with a set of colored patches under three different simulated illumination conditions, while recording brain activity in a 3T fMRI scanner. Our stimuli consisted of a central square patch within a variegated surround. In neutral illumination trials, the surround remained neutrally illuminated while the central patches oscillated between gray and blue or gray and yellow. We used this neutral illumination condition to train a classifier to decode the central color patches into bluish and yellowish colors, based on the activity of the 300 most responsive voxels. These were obtained by means of a localizer that determined the retinotopic location of the target patches. Accuracies were between 50% and 79% correct (mean 67%). We then introduced simulated illumination changes. In these trials, the variegated surround oscillated between a neutral and either a bluish or a yellowish illuminant. This made the neutral gray patches appear to oscillate between gray and yellowish or bluish, respectively. When the classifier, trained on the real colors, was used to decode the appearance of these stimuli, accuracy was slightly reduced but still above chance (46% to 71%, mean 62%). Our results show color appearance can be decoded as early as V1/V2. This indicates that color constancy could be achieved very early in the visual pathways.
Meeting abstract presented at VSS 2016