Marcel Dekker, 1999. — 260 p.This volume is a report of the contributions of a group of us who have been working together in the Arizona State University's Perception Laboratory since 1988. Earlier parts of our work were reported in a volume entitled The Swimmer: An Integrated Computational Model of a Perceptual-Motor System (Uttal, Bradshaw, Dayanand, Lovell, Shepherd, Kakarala, Skifsted, and Tupper, 1992.)
In the years since we finished the manuscript of the first volume describing our SWIMMER project, the work has continued at a rapid pace. Our goal remains the same: to program our model so that it is capable of visually sensing its environment, interpreting that environment within the limits of its own world model, deciding on a course of action, and then behaving in an adaptive manner. As was our original intent, what we have accomplished was not only an engineering tool but also a theory of vision.
In general we have been guided by the same premises on which the original report on the SWIMMER was founded, namely: (I) There is much to be gained by approaching a modeling task from a broader perspective than is usual in this field. Rather than attempt to fine-tune a single-purpose algorithm, we sought to integrate many different procedures and processes into a comprehensive theory of an entire organism. B) From a purely psychobiological point of view, a model that integrates and combines many different weak and idiosyncratic processes into a powerful and capable outcome is a more realistic expression of the way the real organic perceptual-motor system works.
In recent years, these two premises have become more generally accepted, not only as a psychophysical credo but also as a physiological one. Two especially important books have been published that review the substantial amount of information supporting the notion of a separation of visual functions into distinct modular processes as well as different anatomical regions. The first of these is Zeki's (1993) extraordinary review of the specialization of" function in the visual brain. The second, Stein and Meredith's (1993) treatise on the merging of the senses, is not limited to vision alone but deals with the problem of how the different sense modalities interact on both an anatomical and a behavioral level.
There is one grand idea emerging from all of the work in vision science, including ours, these two books in particular, and the efforts of many other scholars, scientists, and engineers in many different fields of biology, psychology, and computer science. That idea is that the brain cum mind is made up of a number of nearly independent channels and specialized centers that deal separately with the different parts of our sensory input.
Our interest in this book is mainly with the visual processes. The decomposable components of an image, as we elaborate in the first chapter, are its attributes, or independently measurable dimensions. The thesis of this book, and of an increasing corpus of modern visual science, is that input information is first analyzed into its component attributes and subsequently synthesized, combined, or (in more modern terminologies) fused or bound into a complete multiattributic perceptual experience. Specifically, the last decade or so has seen an enormous change in the way we believe the organic vision system operates. Vision (with a capital V) now seems to be better described as a collection of "visions" (with a small v).
To make this argument of a collection of "visions," scientists in this field have traditionally used the tools of neurophysiology or neuro-anatomy on the one hand, and of psychophysical and perceptual data on the other. It is obvious, however, as one surveys the literature, that in recent years another research tool has evolved that helps to make this argument. That new tool is the computational model, existing as a series of program steps, operators, and algorithms.
Of course, each of these approaches is limited and incomplete. The leap that is being attempted from the microscopic neuron to the macroscopic perception, for example, is fraught with conceptual and fundamental difficulties as well as with technical ones. The case against an extreme reductionist approach is detailed in an earlier book (Uttal, 1998). There it was argued that neuro-reductionism is computationally impossible and that molar psychophysics is totally incapable of analyzing the constituent mechanisms that underlie the several attributes. Similarly, computational modeling, it must be appreciated, is also limited in what it can accomplish. It is all too often ignored that excellent descriptions are not adequate reductive explanations. Nevertheless, it is clear that none of these approaches is, a priori, superior to any other. Most of the difficulties of one approach are mirrored in those of the others. The triumvirate of modeling, neuroscience, and psychophysical approaches, when used collectively and in mutual support, does, however, create a powerful synergism and produce heuristics that can strongly suggest if not rigorously prove.
Thus, this book merges several traditional approaches to vision. The neuro-physiological, neuroanatomical, and psychophysical literature are reviewed to identify the existing empirical support for the notion of a modular vision system. Then we report the details of our current modular computer model of our "seeing" SWIMMER to elucidate further the plausible and to eliminate the ridiculous.Introduction: A Point of View
Neural and Psychophysical Foundations of a Vision System: Channels and Centers, Interactions and Combinations
Models of Combination and Binding
A Vision System
A Particle System Model for Combining Edge Information from Multiple Segmentation Modules
Combining Images for Three-Dimensional Vision
Object Recognition
Surface Reconstruction
