A good idea about the essential contents of this book of nearly 200 pages (with exclusion of 12 pages of references and an index of 5 pages) may be afforded by the following sentences taken from the text: “We introduce and expose a maximum likelihood framework to be used in computer vision applications when ground truth is available” (page 19). ”...in real cases, the similar images are not identical and therefore when comparing these images a certain distortion between them, called similarity noise, will be present. If one can accurately model the similarity noise distribution, then the retrieval or matching results can be significantly improved by using a suitable metric. The link between the similarity noise distribution and the comparison metric is given by the maximum likelihood theory” (page 18). “We implemented several sophisticated algorithms from the computer vision literature and evaluated their results in the presence of ground truth” (page 17).
Somewhat more analyzing the structure of the book, we can say that it consists of two parts: the theoretical introduction part, consisting of Chapter 1 (Introduction, pp. 1–23) and Chapter 2 (Maximum likelihood framework, pp. 25–59), followed by the application part with the test results on different problem domains in the chapters 3 to 7, with the respective titles: Ch. 3: Color based retrieval (pp. 61–82); Ch. 4: Robust texture analysis (pp. 83–110); Ch. 5: Shape based retrieval (pp. 111–134); Ch. 6: Robust stereo matching and motion tracking (pp. 135–162); Ch. 7: Facial expression recognition (pp. 163–197). Chapter 2 contains what may be called the “mathematical” part of the book, with a description of the maximum likelihood approach and of the experimental setup. In each of the chapters 3 to 7, a typical application has been tested. To a great extent, each of these application chapters starts with some historical overview of the development of the treated subject, followed by a description of the subject itself and of existing algorithms (with ample mention of the references), and finally followed by the conclusions of the experiments. It should be stressed, however, that what we called “the description of the subject” may not be interpreted as a strong (or extended or deep) introduction to that subject. The word “robust” in the title of the book is borrowed from Statistics. In the foreword by S. Huang, member of the editorial board of the series Computational Imaging and Vision, it is said that “an algorithm is robust if it is not very sensitive to departure from the assumptions on which it depends”. A more extended definition of robustness is given by the authors at the beginning of Chapter 2.
A final personal comment of the reviewer concerns the meaning of the word “ground truth”, that rather often has been used by the authors. It is of course possible that either a clear-cut definition or a more vague description has been given in the book, but the reviewer did not find it. In the index, the word is mentioned with references to pages 17 and 50, but at neither place a more explicit description is given. Also reference to the literature is not always one-one; for instance, on page 17 following the word “ground truth”, there is a reference to [Sebe et al., 2000a]. But in the list of references we find Sebe, N. and Lew, M.S. (2000a), but also Sebe, N., Lew, M.S. and Huijsmans, D.P. (2000a), the first one referring to a conference, the second one to a journal. In general, however, the reviewer has the impression that the number of misprints (and such stuff) is not very extended.