Purchase this article with an account.
Edward H. Adelson, Yuanzhen Li, Lavanya Sharan; Image statistics for material perception. Journal of Vision 2004;4(8):123. doi: https://doi.org/10.1167/4.8.123.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
It is impossible to tell the albedo (reflectance) of an isolated Lambertian patch; we must compare it to other, neighboring patches. However, when we look at a more complex material, such as cloth, stucco, or crumpled paper, we can be surprisingly good at judging the albedo, even with isolated viewing. How is this possible? When light strikes a complex surface, it is can be reflected or refracted many times before being absorbed. The pattern of interreflections leads to textural qualities that depend on the albedo and the translucency. In addition, specular reflections are more prominent against dark materials, being diluted when seen on light materials. To quantify the available statistics, we took digital photographs of many materials, and after scaling to a standard mean luminance, we determined the textural properties that could be of use. Among our findings: The power spectrum falls more steeply for light materials than for dark ones. The luminance histogram is quite informative, with the variance, skew, and percentile statistics being diagnostic of surface reflectance properties. Phase coherence across wavelet bands (a measure of structure used by Portilla and Simoncelli's texture algorithm) is diagnostic of translucency. We will discuss a variety of pixel-based and wavelet-based textural measures, and suggest how they may be used in human vision.
This PDF is available to Subscribers Only