Book Reviews Three-Dimensional Computer Vision. by Olivier Faugeras Cambridge, MA: The MIT Press, 1993. Reviewed by: Joseph O'Rourke Department of Computer Science Smith College, Northampton, MA 01063 orourke@cs.smith.edu solved at the end of the book. Despite the wonderful geometry throughout, the author has a predilection for algebraic derivations that might have been curtailed in places. For example, the proposition that "[a] line going through two points...m 1 and m2, is represented by the cross-product m I ^ m2" [p. 15], is proved by noting that a determinant is zero. One could as well view the cross product as the normal to the plane through the camera origin and the two points, which cuts the image plane in the line through them. The notation can grow ugly: Df(x) A x Df(x) T [p. 239], Tr(F(E) F(E) T) [p.255], Skin t (X) = ~ Despite adorning the cover with one of Albrecht Diirer's perspective studies, the figures in this book are drawn without perspective. Fortunately the pedeslrian figures are the weakest aspect of this massive study of geometry-laced vision. Every aspect of this book is generous: its length (663 pages), its margins (5.5cra), the number of bibliographic entries (ksim 300), the 16 pages of figure captions, the long algebraic derivations (four pages to derive the Kalman equations [pp. 308-312]; another four for a kinematic screw equation [pp. 375-379]), the coverage, and the explanatory style. The book starts with an exposition of projective geometry, which is then immediately applied to camera models. Having the projective geometry framework at hand permits discussion of, for example, Buehanan's twisted cubic for camera calibration. This is followed by a thorough coverage of edge detection operators: Roberts, Sobel, Prewitt, Marr-Hildreth, Haralick, Canny, Deriche, Spacek. The last three operators are compared carefully, and the Deriche operator is explored on a real image. After another chapter of geometry, the heart of the book is reached: 3D reconstruction from stereo images, recovery of camera motion in a static environment, and tracking tokens over time. Extensive use is made of the equipolar comtraint: all the possible matches for a particular point in one image of a stereo pair lie on a line, the eqnipolar line, in the other image. The theory is enhanced by frequent reality checks with real images. The sophistication of the presentation in these chapters can be gauged by the simplicity of the proof that there are at most three camera positions compatible with point correspondences, when the points lie on a Maybank quadric: this is deduced by viewing the essential matrices as lying on a projective line in pS, which can intersect the degree-3 manifold of the malrices in at most three points. Optic flow is covered next, primarily specialized to the motion of curves. The focus then shifts to 3D shape representations, with an extensive study of Delaunay triangulations. Again the theory is nicely complemented by polyhedral modeling from real data. The book closes with a discussion of object recognition, with special attention to determining the "pose" of 213 and 3D objects. Nearly every claim is proved in detail, even some not dearly relevant to the main thrust (e.g., ten pages on smoothness properties of the skeleton of a shape [p. 423--433]). Every exercise is [p. 422]. But these are peccadillos. The minor flaws of this impressive tome are lapses of exuberance, and will hardly discomfit a graduate studenL Learning in Embedded Systems by Leslie Pack Kaelbling Cambridge, MA: MIT Press, 1993. 176 pps Reviewed by: Mmrten Boasson Hollandse Sigrmnlapparaten B.V. PO Box 42 7550 GD Hengelo, The Netherlands boasson@ hgl.signaal.nl The book addresses the important problem of adaptation of an embedded system's behaviour to make it more appropriate to the environment in which it is embedded. This process, in its broadest sense, is referred to as learning, and for the author this also includes such simple forms of learning as measuring the width of a hall to allow a measuring robot to perform better. The philosophical debate about what does and does not constitute machine learning is left alone; the interested reader is referred to an earfier publication by the same author. The essential concept in the book is that of action mapping, which maps perceptions from the environment to effector actions. Whereas other methods have been explored, and are carefully referenced, this book focuses on learning action maps directly, without attempting to model the outside world explicitly. The major aim of the book is to present useful algofithm~ for reinforcement learning, i.e., learning by evaluating the state of the environment after each action (c.f. learning by trial and error). The theoretical underpinning of the algorithms is provided when available, but the discussion is not constrained by the existence of theoretical results. The book begins by introducing necessary concepts and simple mathematical foundations. The effect of both faulty sensors and faulty effectors of the system is briefly discussed and some sire- SIGART Bulletin, Vol. 6, No.1

/lp/association-for-computing-machinery/book-review-learning-in-embedded-systems-by-leslie-pack-kaelbling-mit-fTlikDD1Sj