Abstract
Colour constancy stabilises object colours across a large variety of illumination and scene conditions in the natural world. It is unclear whether the neural mechanisms that mediate the phenomenon are optimised for the particular illuminations and surfaces to which we are typically exposed. Classically, colour constancy has been investigated by quantifying the change in appearance of test surfaces with large changes in illumination. Here we measure colour constancy by establishing discrimination thresholds for illumination changes, using a forced-choice paradigm. We compare performance for two different experimental setups: (1) a real, variegated-surface scene inside a viewing box, illuminated by spectrally tuneable multi-channel LED light sources (n=10 participants) and (2) a physics-based computer rendering of a similar scene, displayed stereoscopically on a pair of LCDs (n=9 different participants). For both setups, the illuminations (real or simulated) were metamers from the daylight chromaticity locus or atypical illuminations from an orthogonal locus. On each trial, participants viewed a target illumination (D67), then two subsequent illuminations, one the target illumination and the other a comparison whose chromaticity varied between 1–50 ∆Euv steps from the target. Discrimination thresholds were significantly higher (and therefore, colour constancy was better) for bluer illumination changes along the daylight locus, compared with greener illumination changes on the atypical locus, for both real and simulated scenes. Furthermore, we find no effect on discrimination thresholds of adding 3D objects to the scene, whether they are familiar or novel (fake hand, rendered spheres, painted blocks). Lastly, illumination discrimination thresholds for the real scenes were not significantly different from those for the simulated scenes. We conclude that when assessed via illumination discrimination and when using ∆Euv to parameterise illumination change, the mechanisms of colour constancy are biased for bluish daylight illuminations and that this bias may be measured robustly with real and simulated scenes.
Meeting abstract presented at VSS 2014