|Minimum Probability of Error Image Retrieval|
A visual information retrieval system is a search engine for visual content, i.e. images and video. It accepts a query from a user, searches a visual database, and returns the collection of images that best satisfy the query. Most existing technology is based on pure text-search: the images in the database are annotated with a textual description, the user provides a natural language description of the target image, and retrieval is based on a standard text-based search. Next generation systems will rely on content-based retrieval (CBR), augmenting this text-based search paradigm with one that is grounded directly in the visual domain. Here, the user can also provide the retrieval system with an example image (or image region) and tell the system "the image that I am looking for is similar to this one".
While a large body of work has been devoted to the topic throughout the last decade, there is still a limited understanding regarding the optimality of CBR architectures. This project pursues the development of CBR systems that are optimal in a minimum probability of error sense. This implies the joint optimization of all the components of a CBR system (feature transformation, similarity function, indexing mechanisms, and relevance feedback algorithms) in order to enable queries that are as accurate as possible. The project has originated various contributions at both the practical and theoretical level, and produced a Bayesian retrieval architecture that achieves state-of-the-art results in terms of retrieval accuracy.