Abstract
The ability to discount spatial variation in illumination is central to achieving lightness constancy in natural scenes. There is evidence that the visual system uses the 3D relationship between surfaces as one cue to the likelihood that they share a common illumination (Gilchrist, 1980), but how this cue interacts with the information provided by the luminance statistics of various image regions is not understood. To investigate, we presented subjects with computer-simulated 3D grayscale scenes, using a stereoscopic display. Two planar Mondrians were rendered in a roof-like arrangement. This roof was oriented in 3D such that its ridge formed a horizontal line in the cyclopean image, and each Mondrian appeared tilted away from the observer at a 45° angle. The upper Mondrian contained patches with higher luminances (17 to 262 cd/m2) than the lower one (0.6 to 9.5 cd/m2). On each trial, a target patch was presented at one of four possible tilts. The target could lie flat against either Mondrian, or extend outwards from the horizontal ridge at a tilt that made it coplanar with either Mondrian. Nine target luminances spanned the range 0.6 to 262 cd/m2. Subjects matched the lightness of the target patches to a grayscale Munsell palette that was presented in a separate viewing chamber. The data establish luminance-to-lightness mappings for each target tilt. Both the 3D orientation of the target and its immediate surround in the image affect the luminance-to-lightness mappings: the target appears darker both when its immediate surround is more luminous and when it is co-planar with a more luminous Mondrian. The magnitude of the effect of 3D orientation was about one Munsell step — large enough to be easily measurable, but smaller than has been reported for similar experiments performed using real illuminated surfaces (Gilchrist, 1980).
Meeting abstract presented at VSS 2012