A comparison of global rule induction and HMM approaches on extracting story boundaries in news video

Abstract

This paper presents a multi-modal two-level framework for news story segmentation designed to cope with large news video corpus such as the data used in TREC video retrieval (TRECVID) evaluations. We divide our system into two levels: shot level that assigns one of the pre-defined semantic tags to each input shot; and story level that performs story segmentation based on the output of the shot level and other temporal features. We demonstrate the generality of our framework by employing two machine-learning approaches at the story level. The first approach employs a statistical method called Hidden Markov Models (HMM) whereas the second uses a rule induction technique. We tested both approaches on ∼ 120 hours of news video provided by TRECVID 2003. The results demonstrate that our 2-level machine-learning framework is effective and is adequate to cope with large-scale practical problems.