Retrieval of video story units by Markov entropy rate

In this paper we propose a method to retrieve video stories from a database. Given a sample story unit, i.e., a series of contiguous and semantically related shots, the most similar clips are retrieved and ranked. Similarity is evaluated on the story structures, and it depends on the number of expressed visual concepts and the pattern in which they appear inside the story. Hidden Markov models are used to represent story units, and Markov entropy rate is adopted as a compact index for evaluating structure similarity. The effectiveness of the proposed approach is demonstrated on a large video set from different kinds of programmes, and results are evaluated by a developed prototype system for story unit retrieval.