Geri Grolinger

Impact


Structural Indexing Using Local Image Features

The vision community has only recently begun to apply their recognition algorithms to very large databases. Yet the community has yet to explore the question: For a given feature database size, and a given dimensionality constraint (dictating the complexity of the resulting nearest-neighbour search), how much of the feature's dimensionality should be devoted to local appearance and how much to structural or geometric relations? Moreover, how does the cost of grouping together local features into richer compounds compare with the cost of verifying ambiguous hypotheses? These are essential questions that have received very little attention, and which form the basis of my research.