Abstract

The human visual system is remarkably good at making accurate and reliable interpretations of the world from incomplete or noisy retinal image data. It does this by exploiting implicit knowledge of the statistical regularities of both objects and images. By quantifying the statistical uncertainty inherent in a visual task, ideal observer models specify a limit on the best performance for that task. Comparisons of human and ideal performance, often made by measures of efficiency, can be used to quantify information utilization. Historically, this technique has been developed most fully in the context of visual sensitivity and the early coding of image information. Recently, the ideal observer has been extended to problems of image understanding. I will present an overview of the ideal observer in visual perception research with examples drawn from our work at early and late levels of visual processing. For example, at the image level, human discrimination efficiencies for fractal images are relatively best for fractal dimensions near those of natural image ensembles. At the object level, human classification efficiency for 3D shaded "wire" objects is high enough to exclude generalized 2D template models of recognition and indicates particularly efficient processing of mirror symmetric and coplanar 3D objects.

© 1992 Optical Society of America