The Statistics of Shape, Reflectance, and Lighting in Real-World Scenes

Abstract

2D images are highly ambiguous representations of 3D scenes, and this poses a fundamental obstacle to recovering shape and reflectance from shaded images. A Bayesian approach to overcoming this problem is to exploit statistical regularities in surface shapes, reflectances, and lighting conditions in real world scenes, in order to choose the most likely 3D interpretation of a 2D image. Here I review recent work on the statistical regularities in real world 3D scenes that biological or artificial visual systems could use to overcome image ambiguity, and psychophysical work on the assumptions that the human visual system relies on in order to perceive 3D scenes from 2D images.