Recovering Intrinsic Scene Characteristics from Images

We suggest that an appropriate role of early visual processing is to describe a scene in terms of intrinsic (vertical) characteristics-such as range, orientation, reflectance, and incident illumination-of the surface element visible at each point in the image. Support for this idea comes from three sources: the obvious utility of intrinsic characteristics for higher-level scene analysis; the apparent ability of humans to determine these characteristics, regardless of viewing conditions or familiarity with the scene; and a theoretical argument that such a description is obtainable, by a noncognitive and nonpurposive process, at least, for simple scene domains. The central problem in recovering intrinsic scene characteristics is that the information is confounded in the original light-intensity image: a single intensity value encodes all the characteristics of the corresponding scene point. Recovery depends on exploiting constraints, derived from assumptions about the nature of the scene and the physics of the imaging process. I INTRODUCTION Despite corsiderable progress in recent years, our understanding of the principles underlying visual perception remains primitive. Attempts to construct computer models for the interpretation of arbitrary scenes have resulted in such poor performance, limited range of abilities, and inflexibility that, were it not for the human existence proof, we might have been tempted long ago to conclude that high-performance, general-purpose vision is impossible. On the other hand, attempts to unravel the mystery of human vision, have resulted in a limited understanding of the elementary neurophysiology, and a wealth of phenomenological observations of the total system, but not, as yet, in a cohesive model of how the system functions. The time is right for those in both fields to take a broader view: those in computer vision might do well to look harder at the phenomenology of human vision for clues that might indicate fundamental inadequacies of current aproaches; these concerned with human vision might gain insights by thinking more about what information is sought, and how it might be obtained, from a computational point of view. This position has been strongly advocated for some time by Horn [18-20] and Marr [26-29] at MIT. Current scene analysis systems often use pictorial features, such as regions of uniform intensity, or step charges in intensity, as an initial level of description and then jump directly to descriptions at the level of complete objects. The limitations of this approach are well known [4]: first, region-growing and edge-finding programs are unreliable in extracting the …