Browse Books

Reviewer: Thomas C. Henderson

Robot Vision is an excellent summary of the physicist's approach to computer vision. The central thesis of the book can be stated as follows: The recovery of three-dimensional world information by computer vision can be achieved by understanding the physical interaction of light and surfaces and by using constraints to invert the image formation process. A body of theory related to computer vision, and founded on a strong mathematical basis, is presented. The author attempts to present solutions to basic problems without regard to implementation, and he succeeds at this. However, there are some drawbacks to such an approach. The main drawback would be the inappropriate application of formal techniques to machine vision problems. The author has, for the most part, avoided such pitfalls. The primary purpose of the book is to provide deep coverage of reasonably well-understood topics in computer vision; thus, it is useful as a textbook at the upper-division or graduate level and as a research reference. It nicely complements [1], which covers more topics in less depth. Many of the more elementary concepts of computer vision are not covered in Horn's book, but they can be found, for example, in [2]. Another missing area of discussion is biological vision; however, this material can be found in either [3] or [4]. Finally, there is very limited coverage of pattern recognition techniques; the reader is referred to [5] for more on this subject. The most appealing aspects of the book include the lucid prose, careful artwork, and meaningful exercises. Compared to most other works in this field, the text seems relatively error-free. For people with a strong mathematical proclivity, the text provides problem applications for a variety of theories and techniques. The book does, however, have some failings. First, the title is somewhat inappropriate since the book has almost no discussion of either robots or robotics. There is, moreover, little discussion of vision hardware or implementation. Although purposefully omitted by the author, this seems to be a fundamental philosophical flaw. While it is true that mathematics is essential, one must keep in mind both the computational issues and the synergism of any physical system implementing machine vision. For example, the use of moments is mathematically convenient, but an inappropriate application or implementation of the technique can result in a useless system (e.g., occlusion could prohibit the use of the technique, or a poor implementation could prevent realtime results). Given the current interest in parallel computing, more discussion of these issues would have been highly welcomed. Finally, the book fails to reference several significant works (e.g., books by Fu and Pavlidis), and therefore does not serve as a handy general reference. Robot Vision is what it was intended to be: an exposition of the extant strong mathematical results in computer vision. It is well written and contains a wealth of information collected together in one volume. As such, it is indispensable to any serious student or researcher in computer vision.