Shape-based Illustration Indexing and Retrieval Some First Steps

We propose a general set of ideas for indexing technical illustrations based on the shapes present in them, so that they can be e ciently retrieved later using as the key other `similar-looking' illustrations (either pre-existing, or interactively drawn by the user). A very simple prototype system demonstrating these ideas was implemented and is described in this note. The general scheme is to select a class of basic shapes and record in the index where (more precisely, via what homothetic map) these basic shapes match well into the illustration. The current implementation indexes using line segments as the only basic shape. A Hausdor matcher is then used to compute the best alignment of the basic shape matches between a query and an illustration, thus giving us a measure of the distance between the two. Currently every illustration in the data-base is matched individually against the query, though sublinear algorithms are under investigation. A library of approximately two-hundred illustrations from a geometry textbook was indexed using this scheme and then used for retrieval experiments. An interactive interface was provided for specifying the data-base to be searched and the query illustration, for setting various parameters regarding the match, and for displaying the best matches found in the data-base.