Geri Grolinger


Structural Indexing Using Local Image Features

We are planning to develop a method for combining two or three SIFT features into a single compound feature. In addition to retaining the encoding of each patch, we will encode the relations between the features, including relative scale, orientation, and proximity. Clearly, adding additional features to an index, along with a set of relations, will increase the dimensionality of the compound feature over the original SIFT feature which, in turn, will increase the complexity of the nearest-neighbour search algorithm. Therefore, we will reduce the dimensionality of the compound feature down to that of the original SIFT feature. By combining two or three SIFT features, we expect indexing ambiguity to decline as the index becomes richer and more specific. On the other hand, the process of combining the features into meaningful groups of two or three comes at a certain computational cost. Studying this tradeoff will help us find an optimal balance between bottom-up feature extraction and top-down model recovery as a function of database size.