Publisher's Synopsis
The main issues of this book concern the way that knowledge about the real world (geometry of surfaces and the projection of a 3D screen into a 2D picture, etc) can provide strong constraints on the possible interpretation of a picture; how apparently intractable computational problems may have interesting subproblems which are tractable; and how a provably-correct and efficient algorithm can provide a solid foundation on which to build a robust and practical computer vision system.;The book is divided into two parts. Part 1 describes how to use general non-model-based methods, such as line labelling, to gain 3D information about scenes, while at the same time solving the correspondence problem between pairs of pictures in order to learn the 3D shape of objects.;Part 2 looks at how to interpret a picture of a cluttered scene containing known objects. Efficient algorithms are described for the problem of determining all physically possible interpretations of such a picture. Computational complexity and ambiguity are discussed in detail. Although originally designed for two-dimensional objects viewed under ortho-graphic projection, the algorithms can also be applied to the cases of perspective projection and range imagery. An extensive survey of the literature (over 100 references) covers all presently-used techniques for the model-based recognition of partially-visible objects.
