Spatial ratios of cone excitations from natural scenes over the course of the day

The light reflected from natural scenes varies markedly during the day. In addition to spectral changes in the illumination on a scene from the sun and sky, the geometry of the illumination alters with the sun's elevation and cloud cover and with changes in local surface shadow, mutual reflection, and haze. One low-level visual signal that offers an invariance to spectral changes in illumination is the spatial ratio of cone photoreceptor excitations generated in response to light reflected from points or surfaces in a scene. Such a signal may indeed provide a basis for several kinds of color judgments. But it is not obvious that these ratios are also invariant under geometric changes in illumination. The question of this invariance was addressed here by analyzing sequences of time-lapse hyperspectral radiance images from five outdoor rural and urban scenes. The images were acquired with a hyperspectral camera with spatial resolution 1344 x 1024 pixels and line-spread function that was approximately Gaussian with standard deviation 1.3 pixels at 550 nm. In each scene, 10,000 pairs of points were sampled randomly and for each pair the deviation in spatial cone-excitation ratios were estimated at successive times of day. Averaged over scenes, time intervals, spacings of the pairs of points, and cone classes, the mean relative deviation in ratios (i.e. the symmetric mean absolute percentage error) was more than 20%, much greater than the 4% found with pure spectral changes in illumination. But when time intervals were restricted to about an hour or less and pair spacings to 4 pixels or less, the mean relative deviation fell to 7% and for four of the five scenes to 5%. For cone signals adjacent in time and space, spatial ratios may provide a reliable cue for making visual judgements even in a varying natural environment. (presented at VSS 2015, Tradewinds Island Resorts, FL, May 15-20, 2015)