Perceptual grouping is the phenomenon of putting items in the visual field together, or of "organizing" image data. It is an "intermediate" level process in vision. Such a process of organizing image data requires the use of rules and constraints defined in terms of the image plane entities, such as the Gestalt laws of organization. These constaints are those that are invariant under the projection process. The problems of perceptual grouping in the context of computer vision then are (i) To identify important intrinsic properties of the image plane tokens and important image plane relationships among them, and (ii) To devise a computational framework in which these properties and their image plane relationships can be identified and larger units can be constructed from these tokens and groups of tokens. The rules to detect the organization need to be completely stated in terms of intrinsic properties of tokens being grouped and their image plane relationships. This research is aimed at studying the grouping phenomenon from a computational viewpoint for patterns containing "generalized" tokens. These tokens can be a subset of dots, line segments, curves, and globular or elongated blobs with arbitrary shapes. Our approach primarily involves the use of the Voronoi tessellation (and its dual Delaunay graph) as the underlying geometric representation of the spatial distribution of tokens in the scene. A grouping built upon this representation will enforce non-local Gestalt constraints such as alignment of tokens along smooth curves to resolve ambiguous situations. This grouping process integrates the information from different sources each of which detects groupings independently based on token properties such as orientation or size. The algorithms developed in this project will be tested on real images and the usefulness of these grouping algorithms will be assessed in such areas as gap filling in object boundaries, segmenting homogeneous textured regions, and computing optical flow in dynamic images.
