Summarizing the Visual Via the Verbal
Manual indexing and video annotation, as well as video summarization can be costly and time consuming, taking into account the increasing amount and diversity of multimedia information provided by multimedia production industries. While most automatic film summarization approaches focus on content-based video retrieval, this paper presents a novel method, using different kinds of collateral texts. Such texts describing film content can be freely downloaded on the Internet and provided by organisations producing audio description. A multi-disciplinary approach is suggested, combining cross-document coreference and information extraction techniques, to automatically produce film summaries. The method is inspired by the lexical analysis of a corpus of plot summaries, including short overviews of the film story, and a corpus of audio description, including time-coded detailed descriptions by experts for visually impaired people. The preliminary user evaluation of the method shows encouraging results regarding the precision and ranking of the retrieved video shots. The method may be adapted for different kinds of data and evaluated it in different contexts, such as virtual meeting summarization and browsing.
Keywords: Cross-Document Coreference, Video Summarization, Collateral Text, Information Extraction
Dr. Eleftheria Tomadaki
The Open University