Adam Reeves¹; ¹Northeastern University

As everyone knows, the environment appear ‘rosy’, or red-tinged, at dawn, even after full adaptation to the skylight. This alone demonstrates that colors are not constant, even if there is a tendency towards constancy. To capture this, two metrics are the constancy index CI=1−b/a and the Brunswick Ratio BR =c cos()/a, where a, b, and c are sides of the color constancy triangle. Ideally, CI=BR=0 for no constancy and CI=BR=1 for perfect constancy. In the canonical Arend et al. (1991) study, LA made matches to 5 hues, each illuminated by an evening 4,000oK sky, a mid-day, bluish sky of 10,000oK, and a neutral sky of 6500oK. His mean CI, 78%, was close to the mean over 60 studies of 74% reported by Foster (2011). However, LA’s mean BR was 95%, a gross exaggeration as his matches for 4 of the 5 hues fell well short. This problem is not unique to LA but typical. A further problem is that BR is indiscriminant: if the subject picks matches at random from RGB color space, the BR given (e.g.) a simulated 4,000 oK illuminant ranges across hues from +3.1 to -2.3, including +1.03 (‘perfect color constancy’) for a u’=.20, v’= .46 test patch. If the subject picks a single primary (R, G, or B) on every trial, BR ranges from −12 to +15, again straddling 1.00. In contrast, CI is always negative whether the subject picks at random from all of RGB space, from a limited region around white, or from any one spot, making it easy to filter data for bad responses. Recent papers have increased reported constancy towards 100%, but this is an artifact of employing BR, as can be easily seen by looking at the raw data. As a descriptor, CI is clearly superior to BR.

Acknowledgements: No funding, but inspiration from D. H. Foster