Indexing and object recognition using symmetry, invariants and quasi-invariants

This thesis concerns the general problem of three dimensional model-based vision. Our focus in this research is on the use of symmetry, invariants and quasi-invariants as cues to determine object identity and pose from a single perspective image, given a database of geometric models of objects. A fundamental difficulty in recognizing objects from images is that the appearance of an object is viewpoint dependent. Matching between an image and an object can involve search through a high dimensional space of possible viewpoints. This is a very inefficient process. We seek to improve the matching performance by utilizing symmetry, invariants and quasi-invariants to recover the orientation of an object from detected symmetric features. Another major unsolved problem in object recognition is the construction of efficient and robust indexing functions for large model databases. In general, a good indexing function should directly identify a model in the database using information recovered from an image. We investigate the construction of such indexing functions based on invariant properties of the projections of object contours. Quasi-invariant properties are used to verify the generated hypotheses. Using invariant and quasi-invariant properties, we have developed and implemented new efficient techniques for recognizing and locating flat symmetric objects, polyhedra with planes of symmetry, and straight homogeneous generalized cylinders from a single perspective image. In addition, a method which takes uncertainty propagation into account is used to improve the accuracy and efficiency of our recognition system. Results are demonstrated on a wide variety of real and synthetic images