Browse Books

Reviewer: Sebastien Lefevre

After several decades of research, computer vision is still a challenging domain. Despite its continuous improvement, the computer is far behind the human in understanding visual content. Fortunately, the computer vision research community is very active, with contributions from many domains, especially computer science, applied mathematics, and physics. Thus, it is no surprise that new solutions often come with a deep mathematical background. At the same time, computer vision applications are becoming more and more common in today's life. This is yet another new book in computer vision-the field benefits from many Web surveys and introductory texts. But since the field is very active, new coverage is needed on a regular basis. Thus, one of the main interests of Richard Szeliski's book is to give a rather up-to-date overview of the state of the art. Moreover, it also comes with plenty of exercises (unfortunately, not always self-contained) and project proposals, which make this book a valuable resource for teaching computer vision at either the undergraduate or graduate level. Focus is given to techniques that actually solve real-world problems rather than solutions with nice mathematical formulations. The book is thus easy to read, and most of the computer vision problems are presented in a comprehensive way. Problems are also addressed (when possible) from a consumer's point of view; this way, readers can apply computer vision techniques to their own data, and project proposals look much more appealing. However, if this book is studied with the aim of designing a computer vision system, the reader might be quite disappointed when looking for algorithms since no code is given. The book covers computer vision quite exhaustively, addressing 13 topics. The first four are related to digital image processing and discuss image formation, image filtering, feature detection, and image segmentation. Image processing is indeed very close to computer vision, even if this is not explicitly stated enough in the book. From an image processing point of view, the book's content is quite disappointing. For instance, color models (for example, those based on the human vision system) receive insufficient coverage despite their importance to computer vision. However, it is probably the description of mathematical morphology, presented here as a filtering toolbox with a few of its prehistoric operators, that should definitely be avoided if one would like to have a correct introduction to this valuable theory in image processing. The rest of the book covers many computer vision problems, including: how to deal with pairs of images or features in order to achieve registration or stereo correspondence; how to extract motion and use it to recover structure or build mosaic images; and how to build a visual representation using 3D reconstruction and rendering. Machine learning is also partly addressed in the chapter dedicated to recognition. The first two appendices recall some of the necessary mathematics since, as previously stated, most computer vision techniques were designed on strong mathematical foundations. A valuable resource is the last appendix. It includes a list of datasets, software, slides and lectures, and references for further reading. However, the author's tendency to self-reference may disturb readers looking for appropriate material for further reading in the bibliography. Despite the criticism, the book is an interesting read for any student or engineer who wants a broad introduction to the field of computer vision. Then, readers can further explore the topic with readings from, for instance, the Annotated Computer Vision Bibliography [1]. From a teaching point of view, the book is a valuable resource, offering an extended list of exercises, project proposals, and appealing applications of computer vision techniques. Online Computing Reviews Service