Towards an all-purpose, content-based multimedia information retrieval system (Master Thesis, Finished)
Digital media in its various forms is ubiquitous and important in many different domains ranging from culture and arts to science and, of course, our everyday lives. As our media-collections grow at an ever-increasing pace - not only in terms of sheer volume but also in terms of variety - it becomes more challenging to manage those collections and make effective use of the knowledge they contain. One of the major obstacles is retrieving an item of interest from such a collection. Traditionally, this was facilitated by keyword-based search, which requires prior, manual annotation of the data. Therefore, it comes as no surprise that this technique scales badly as the velocity of data-generation increases.
Content-based multimedia retrieval (CBMR) aims at offering some relieve here. As opposed to the classical keyword search, content-based retrieval techniques directly leverage the content of the media files. Thus, no manual annotation is required, even though it can still be used to complement CBMR methods. Typical query modes include Query-by-Example or Query-by-Sketch. Over the past decades, umpteen approaches for the different modalities have been developed, refined, implemented and evaluated and a lot of best practices have evolved.
This thesis explores and evaluates different, media type-specific content-based multimedia retrieval techniques and integrates them into a single system with the aim to devise a solution that can manage and search mixed multimedia collections. The work is based on the existing vitrivr system, an open source full-stack content-based multimedia retrieval system with focus on video. Vitrivr's modular architecture makes it extendable and that architecture will be leveraged to add support for still images, audio and simple 3D models in addition to its existing video retrieval capabilities.
Start / End Dates
2017/01/09 - 2017/07/08