Similarity Search with Multiple-Object Queries

Richard Connor, Alan Dearle, David Morrison, Edgar Chávez

Abstract: Within the topic of similarity search, all work we know assumes that search is based on a dissimilarity space, where a query comprises a single object in the space. Here, we examine the possibility of a multiple-object query. There are at least three circumstances where this is useful. First, a user may be seeking results that are more specific than can be captured by a single query object. For example a query image of a yellow hot-air balloon may return other round, yellow objects, and could be specialised by a query using several hot-air balloon images. Secondly, a user may be seeking results that are more general than can be captured by a single query. For example a query image of a Siamese cat may return only other Siamese cats, and could be generalised by a query using several cats of different types. Finally, a user may be seeking objects that are in more than a single class. For example, for a user seeking images containing both hot-air balloons and cats, a query could comprise a set of images each of which contains one or other of these items, in the hope that the results will contain both. We give an analysis of some different mathematical frameworks which capture the essence of these situations, along with some practical examples in each framework. We report some significant success, but also a number of interesting and unresolved issues. To exemplify the concepts, we restrict our treatment to image embeddings, as they are highly available and the outcomes are visually evident. However the underlying concepts transfer to general search, independent of this domain.