Nowadays, the similarity search in multimedia databases is performed through similarity queries explicitly specified by users. The queries return a certain part of the database that is relevant to the user specified query parameters. However, this approach suffers in case the user does not know how to specify the query, or actually she/he only wants to know what the database contains in the whole picture. In such case non-standard access to data is more appropriate, e.g., the exploration of a multimedia database.
During the exploration process the user gains a complex idea of all the stored data rather than a particular part of database returned as the result of some similarity query. In the complex view the user is osupported in browsing the space of multimedia data (typically by multi-touch device, provided by modern technologies, e.g., iPad) and that results in stream of similarity queries. For a convenient user-friendly browsing, the exploration system has to evaluate these queries promptly, which is not guaranteed in case of standard query processing (even approximate). Hence, the goal of this project is to propose and implement access methods that provide functionality of real-time similarity retrieval, thus founding fundamentals for user-friendly exploration of multimedia databases.

The task of smart similarity search in huge multimedia databases remains still a big challenge for the image retrieval research. The domain experts try to design more effective retrieval models which often utilize complex and expensive similarity measures. At the same time, the database experts developing similarity-based indexes have to fight with less "indexable" similarity spaces induced by the more complex similarity measures. Moreover, since there is a common practice that the provided similarity measures are considered as black-box algorithms, only general ap-proaches enabling efficiency tuning can be employed (e.g., the TriGen algorithm). However, the general approaches are not sufficient and so the database experts can no longer stand aside from the similarity modeling. In this project, we would like to "open the Pandora's box" and enter the world of domain experts, in order to design "indexable" similarity spaces. In general, we would like to describe a new similarity space modeling schema considering not only the retrieval quality but also the efficiency issues. In other words, we plan to introduce indexability measures to the similarity modeling process and investigate more variants of popular multimedia distance spaces. As we have recently shown [Beecks et al. 2011] in cooperation with RWTH Aachen, this ap-proach can result in interesting tradeoffs, where at least two orders of magnitude speedup can be achieved for the price of only slightly decreased retrieval quality.

The volume of unstructured databases grows extremely whereas its annotation is problematic. The similarity search concept based on a similarity function defined for each pair of database objects is more suitable for this kind of data. The similarity is usually modelled by a distance function satisfying metric axioms, which allows efficient indexing. However, metric axioms can be very restrictive for domain experts who may prefer non-metric functions. Hence database experts have to solve this problem by converting non-metric functions to metric ones or by developing new types of non-metric indexing methods.

One of the areas where metric/non-metric similarity searching is used is computational proteomics. During the determination of the biological function of an "unknown" protein, retrieval of "known" proteins with similar structures (and thus probably with similar function) is very useful. Moreover, fast and cheap determination of protein structures is also an open problem. From database point of view, it is possible to use databases of known protein structures to address this problem. In this approach, sequence-structure similarity functions are used to obtain structures that can be similar to the searched structure of the protein.

Our goal is developing high-quality structure and sequence-structure similarity functions and methods for their indexing.