The role of saliency and error propagation in visual object recognition
T.D. Alter, W. Eric L. Grimson · Medical Entomology and Zoology · 1995
Man-made environments typically contain many objects with well-defined boundary edges, such as staplers, tables, and computers. These edges may determine the shapes of the objects and are invaluable in their identification. This thesis is about using such edge information for recognizing three-dimensional objects in images. In particular, we consider an approach that first identifies image features that are likely to come from a single object, and later uses these features to project similar model features into the image for verification. In light of this approach, we focus on two critical problems: identifying salient curves in images and the propagation of error from image features to projected model features. In our study of saliency, we analyze one of the notable approaches to the problem, Shashua and Ullman's Saliency Network, and propose a new method based on shortest-paths algorithms. The new method is designed to overcome some of the most critical problems in the network. Using shortest-paths methods, we find the optimal curves in the image according to a measure that prefers contours that are long, smooth, and connected. For the verification stage, we consider how to make the verification robust to locational errors in the image features. Error in the image features leads to uncertainty in the projected model features, which deters effective recognition. We show how error propagates when poses are based on three pairs of model and image points, for both Gaussian and bounded error in the detection of image points, and for both scaled-orthographic and perspective projection models. This result applies to objects that are fully three-dimensional, where past results considered only two-dimensional objects. In addition, we show how we can utilize linear programming to compute the propagated error region for any number of initial matches. Finally, we use these results to extend, from two-dimensional to three-dimensional objects, robust implementations of alignment, interpretation-tree search, and transformation clustering. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)