ERCIM News No.47, October 2001 [contents]
MIA - Querying Multimedia Databases
by Arjen de Vries
The Multimedia Information & Analysis (MIA) project investigates the problem of access to multimedial collections.
Finding textual information on the ever growing Web is already quite a challenge, let alone searching the forthcoming wealth of multimedia documents. This research provides the basis for a new generation of search engines on a multimedia Web, and supports digital archiving in large companies. It is carried out by CWI and seven partners (universities as well as companies) with funds allocated to the Amsterdam Science and Technology Centre (WTCW).
Digital multimedia documents are best stored centrally and distributed through broadband Internet, allowing remote access from anywhere anytime; shortly, even from your mobile phone. And, digital data are easily exchanged, also enabling the reuse of existing material. Furthermore, people at home create ever more digitized data: making your own compact discs with music has become extremely popular, and digital photo- and film-cameras become increasingly common. Technology developed in MIA focuses therefore on retrieving that song with that nice little melody, or that picture of our sailing, with the Boston skyline; when the weather was nice, was it 1997, or 1998?
MIAs database team at CWI focuses on two themes: scalability and query formulation. The target in the first theme is processing queries in a collection exceeding one million pictures. Common approaches based on the colour distribution of an example image break down for collections of this size, because often either all objects or none are found. Moreover, some back of the envelope calculations show that the required time per iteration of the search process yields unacceptable response times as soon as the collection grows beyond several thousands of pictures. Yet, the toughest problem in searching picture archives with current systems, is that query results are hard to understand even for their software developers. The end-user cannot be expected to comprehend why an example picture of a sunset at sea results in (however beautiful) pictures of African savannas. Such inaccurate answers are an unwanted side-effect of the uncertainty inherent to posing the question, exposing the wide gap between our high-level perception and the (necessarily) much simpler methods for image analysis.
We seek to reduce this gap with better formulated search requests, based on two different but complementary approaches. On the one hand, we assist the user with automatic query formulation; in our opinion, many improvements can be achieved by collecting more information about the users interest in a carefully crafted, interactive query process. On the other hand, we attempt to give users more insight into the query process inside the system itself, enabling them to intervene in that process and eventually adjust it (query articulation). An interesting challenge in our research is to find the correct balance between these two approaches.