AT&TV is an information retrieval system for broadcast television and radio developed as part of the DART project. Unlike some other systems, it does not rely on manual classification or annotation of the broadcast material--the material is indexed automatically from the air. While many digital video library projects focus on broadcast news, we have concentrated our efforts on large-scale indexing of general TV output to produce a scalable and usable system.
We have investigated the use of both video and audio stream analysis, together with closed-caption and speech recognition derived text transcripts, to aid retrieval of whole programmes and of segments within them. Our initial system used speech recognition to generate transcripts for indexing of radio and television news broadcasts. The current system uses closed-caption transcripts where available, to index a much wider range of television programmes. A centralised archive is maintained of the past seven days television across the four main UK terrestrial channels, thus removing a viewer's need to select programmes prior to broadcast. Users can simply type in queries and watch any matching programmes or parts thereof. The system delivers this tailored video-on-demand to clients with broadband access, and is in regular use in a lab-wide deployment with around 50 users.
In conjunction with the current system, we have also implemented a simple user interest profile system. This allows users to register general interests in the form of AltaVista-style queries (e.g. 'golf', '+music +blues', '+films -"jeff goldblum"', etc.) after which the user will be notified automatically via email of any newly indexed programmes which contain segments matching their interests. We also use Local Context Analysis of the archived programme transcripts to provide an intelligent query-expansion recommendation facility, allowing users to redefine their interests to make the retrieved programmes more suitable.
ScreenshotsTo get a feel for how AT&TV works, take a look at the following screenshots of the system in action.
As with Shoebox, underlying the user-level features of AT&TV there is an ODMG-compliant object-oriented database (OODB) which was designed and developed here at AT&T Laboratories Cambridge, primarily as a vehicle for experimentation in the field of multimedia information retrieval. Because AT&TV is based on this heavyweight database, it will easily scale to handle very large collections of video.
Copyright © 2002 AT&T Laboratories Cambridge