Colin C. Hand
This thesis is concerned with indexing in model-based computer vision. The central problem addressed is: given a database of object models known to a system, how can one avoid exhaustive search when matching those models with a particular image? The problem of attempting to index only a limited subset of the database of models for a full model match has received little attention. This thesis introduces a new indexing technique using indexing features (i.e. subsets of various object's full descriptions) which map onto different object models under a non-similarity transformation. This allows a feature to simultaneously model both similarities and differences holding between objects; the same indexing feature maps onto phenomenally different forms. Different objects related to a feature will have phenomenally instantiations contained within them. Therefore, one can index those objects by determining the parameter values of the non-similarity transformation between the feature and an image instance of some object. Once a model has been indexed by an indexing feature this hypothesis can be verified by a full match with the image. The non-similarity transformation investigated is anisotropic scaling in the image alone. It has been implemented via the Generalised Hough Transform.
This paper is not available online