Generalized Shape Autocorrelation

This paper presents an efficient and homogeneous paradigm for automatic acquisition and recognition of nonparametric shapes. Acquisition time varies from linear to cubic in the number of object features. Recognition time is linear to cubic in the number of features in the image and grows slowly with the number of stored models. Nonparametric shape representation is achieved by spatial autocorrelation transforms. Both acquisition and recognition are two-step processes. In the first phase, spatialautocorrelationoperators are applied to the image data to perform local shape analysis. Then, spatial autocorrelation operators are applied to the local shape descriptors to either create entries (acquisition) or index (recognition) into a table containing the distributed shape information. The output of the table is used to generate a density function on the space of possible shapes with peaks corresponding to high confidence in the presence of a particular shape instance. The behavior of the system on a set of complex shapes is shown with respect to occlusion, geometric transformation, and cluttered scenes.