List of Past Events
Expanding The Domain of Color Constancy
Qasim Zaidi
Monday, November 08, 2004, 02:00pm - 03:00pm
SUNY College of Optometry, VIsion Sciences
Identification of objects and materials across illuminants is one of the most important functions of color vision. The unresolved question is whether identification is based on constancy of perceived appearance, or on inference of spectral reflectance. I will begin by describing the invariants of the natural world that are critical to color-based identification. These invariants are based not on statistical distributions of spectral reflectance or illuminant spectra, but on strong correlations between cone-absorptions across illuminants. Because of these correlations, an adaptation mechanism that shifted every illuminant�s cone-coordinates to that of a reference illuminant would lead to approximately constant cone-coordinates for spectral reflectances. Under conditions of prolonged adaptation to a single illuminant, a color categorization task reveals that appearance is almost constant across illuminants. Consequently, an observer can use similarity between perceived colors to make accurate identifications. I will show that, contrary to general belief, this adaptation is spatially local and temporally extended. In more natural situations, more than one illuminant is present on a scene, and appearances vary across illuminants. The invariant color correlations can also be exploited in non-adaptive algorithms to simultaneously estimate the shift in illuminant cone-coordinates and identify objects across illuminants. In direct tests, observers make generally accurate identifications despite appearance inconstancy. The few errors that observers make are systematic and indicate an identification strategy that compares color differences between objects to color differences between backgrounds. The reliable use of similarities between color differences expands the domain of color constancy to a wide range of natural circumstances.
