Abstract

A major goal of shape analysis is to derive discrete qualitative descriptions of binary data which may assume a continuum of distributions in space—witness the range of 2-D silhouettes that may be called hammer. We address the question, when during information processing should a representation proceed from a continuous, iconic data structure (i.e., in the format of the image array) to more abstract, discrete symbolic descriptors? Conventional approaches to shape representation immediately describe shape in terms of the spatial relationships among instantiations of a small number of shape primitives. Depending on the primitives chosen and the rules for assigning primitives to portions of the image, this strategy can lead to unnatural and unstable descriptions. We argue that at an early level of shape analysis the quantity of interest is the continuous-valued location and density of matter in space, across all scales of resolution. A natural data structure is scale-space, the original signal smoothed at all scales.¹ Shape information is manipulated via operations for moving matter around in the scale-space image. Different shapes can be compared through comparison and alignment of their scale-space images under these operations.

© 1985 Optical Society of America