Stefano Soatto – författare

This book gives senior undergraduate and beginning graduate students and researchers in computer vision, applied mathematics, computer graphics, and robotics a self-contained introduction to the geometry of 3D vision; that is the reconstruction of 3D models of objects from a collection of 2D images. Following a brief introduction, Part I provides background materials for the rest of the book. The two fundamental transformations, namely rigid body motion and perspective projection are introduced and image formation and feature extraction discussed. Part II covers the classic theory of two view geometry based on the so-called epipolar constraint. Part III shows that a more proper tool for studying the geometry of multiple views is the so- called rank considtion on the multiple view matrix. Part IV develops practical reconstruction algorithms step by step as well as discusses possible extensions of the theory. Exercises are provided at the end of each chapter. Software for examples and algorithms are available on the author's website.

This book is intended to give students at the advanced undergraduate or introduc­ tory graduate level, and researchers in computer vision, robotics and computer graphics, a self-contained introduction to the geometry of three-dimensional (3- D) vision. This is the study of the reconstruction of 3-D models of objects from a collection of 2-D images. An essential prerequisite for this book is a course in linear algebra at the advanced undergraduate level. Background knowledge in rigid-body motion, estimation and optimization will certainly improve the reader''s appreciation of the material but is not critical since the first few chapters and the appendices provide a review and summary of basic notions and results on these topics. Our motivation Research monographs and books on geometric approaches to computer vision have been published recently in two batches: The first was in the mid 1990s with books on the geometry of two views, see e. g. [Faugeras, 1993, Kanatani, 1993b, Maybank, 1993, Weng et aI. , 1993b]. The second was more recent with books fo­ cusing on the geometry of multiple views, see e. g. [Hartley and Zisserman, 2000] and [Faugeras and Luong, 2001] as well as a more comprehensive book on computer vision [Forsyth and Ponce, 2002]. We felt that the time was ripe for synthesizing the material in a unified framework so as to provide a self-contained exposition of this subject, which can be used both for pedagogical purposes and by practitioners interested in this field.

This book is intended to give students at the advanced undergraduate or introduc­ tory graduate level, and researchers in computer vision, robotics and computer graphics, a self-contained introduction to the geometry of three-dimensional (3- D) vision.

Images contain information about the spatial properties of the scene they depict. When coupled with suitable assumptions, images can be used to infer thr- dimensional information. For instance, if the scene contains objects made with homogeneous material, such as marble, variations in image intensity can be - sociated with variations in shape, and hence the “shading” in the image can be exploited to infer the “shape” of the scene (shape from shading). Similarly, if the scene contains (statistically) regular structures, variations in image intensity can be used to infer shape (shape from textures). Shading, texture, cast shadows, - cluding boundaries are all “cues” that can be exploited to infer spatial properties of the scene from a single image, when the underlying assumptions are sat- ?ed. In addition, one can obtain spatial cues from multiple images of the same scene taken with changing conditions. For instance, changes in the image due to a moving light source are used in “photometric stereo,” changes in the image due to changes in the position of the cameras are used in “stereo,” “structure from motion,” and “motion blur. ” Finally, changes in the image due to changes in the geometry of the camera are used in “shape from defocus. ” In this book, we will concentrate on the latter two approaches, motion blur and defocus, which are referred to collectively as “accommodation cues.

Images contain information about the spatial properties of the scene they depict. When coupled with suitable assumptions, images can be used to infer thr- dimensional information. For instance, if the scene contains objects made with homogeneous material, such as marble, variations in image intensity can be - sociated with variations in shape, and hence the “shading” in the image can be exploited to infer the “shape” of the scene (shape from shading). Similarly, if the scene contains (statistically) regular structures, variations in image intensity can be used to infer shape (shape from textures). Shading, texture, cast shadows, - cluding boundaries are all “cues” that can be exploited to infer spatial properties of the scene from a single image, when the underlying assumptions are sat- ?ed. In addition, one can obtain spatial cues from multiple images of the same scene taken with changing conditions. For instance, changes in the image due to a moving light source are used in “photometric stereo,” changes in the image due to changes in the position of the cameras are used in “stereo,” “structure from motion,” and “motion blur. ” Finally, changes in the image due to changes in the geometry of the camera are used in “shape from defocus. ” In this book, we will concentrate on the latter two approaches, motion blur and defocus, which are referred to collectively as “accommodation cues.