Abstract

In the shape-from-shading world, one typically assumes that surfaces have uniform reflectance and that all luminance changes are due to changes in illuminance. In the Mondrian world, used in the study of lightness and color constancy, one assumes that the illuminance is nearly uniform and that all sharp changes in luminance are due to reflectance. But in the real world, luminance variation arises from both reflectance and shading. We are interested in segregating image luminance into its intrinsic sources, illuminance and reflectance, in a manner consistent with human perception. For certain simple images, small manipulations of the 2-D configuration can lead to large changes in the perceived 3-D structure; viewers report concomitant large changes in the perception of reflectance and shading under these manipulations. We have developed an algorithm that decomposes images of some simple polyhedral objects into appropriate intrinsic images and captures some of the qualities of the way humans perceive these images. The algorithm contains a set of interacting subprocesses, each of which seeks to explain as much of the image data as possible, thereby minimizing an overall cost function. A shape process infers a 3-D shape from contours of the observed 2-D image, a shading process seeks the light source position that accounts for as much of the data as possible in terms of illuminance variation, and a reflectance process takes care of any image data that cannot be explained by the 3-D model coupled with shading.

© 1989 Optical Society of America