Database Support for Large-Scale Multimedia Retrieval (PhD Thesis, finished)
With the increasing proliferation of recording devices and the resulting abundance of multimedia data available nowadays, searching and managing these ever-growing collections becomes more and more difficult. In order to support retrieval tasks within large multimedia collections, not only the sheer size, but also the complexity of data and their associated metadata pose great challenges, in particular from a data management perspective. Conventional approaches to address this task have been shown to have only limited success, particularly due to the lack of support for the given data and the required query paradigms. In the area of multimedia research, the missing support for efficiently and effectively managing multimedia data and metadata has recently been recognised as a stumbling block that constraints further developments in the field.
In this thesis, we bridge the gap between the database and the multimedia retrieval research areas. We approach the problem of providing a data management system geared towards large collections of multimedia data and the corresponding query paradigms. To this end, we identify the necessary building-blocks for a multimedia data management system which adopts the relational data model and the vector-space model. In essence, we make the following main contributions towards a holistic model of a database system for multimedia data: We introduce an architectural model describing a data management system for multimedia data from a system architecture perspective. We further present a data model which supports the storage of multimedia data and the corresponding metadata, and provides similarity-based search operations. This thesis describes an extensive query model for a very broad range of different query paradigms specifying both logical and executional aspects of a query. Moreover, we consider the efficiency and scalability of the system in a distribution and a storage model, and provide a large and diverse set of index structures for high-dimensional data coming from the vector-space model.
The developed models crystallise into the scalable multimedia data management system ADAMpro which has been implemented within the iMotion/vitrivr retrieval stack. We quantitatively evaluate our concepts on collections that exceed the current state of the art. The results underline the benefits of our approach and assist in understanding the role of the introduced concepts. Moreover, the findings provide important implications for future research in the field of multimedia data management.
- Blog article on ADAMpro from 05.09.2014: Big data tamed with the cloud (Microsoft Research Blog)
- GitHub repository, Docker repository
- Full text available at edoc
Prof. Dr. Michael Grossniklaus, Universität Konstanz
Date of Defense
Date of Publication
CHIST-ERA / SNF, Project IMOTION