Performance evaluation of media segmentation heuristics using non-Markovian multi-class arrival processes

In the current scenario of a multitude of digital audiovisual sources it is valuable to set up systems capable to automatically analyze, classify and index the material for further usage. In this paper we propose a technique to study the performance of a system for the automatic segmentation of a particular kind of television program: television news.In the analyzed system, the segmentation is performed thanks to a set of heuristics that have to be tailored for the particular program structure they are working on.We model the bulletin broadcasts as non-Markovian multi-class arrival processes and we generate newscasts as their constituting parts. We exploit this model to simulate and study the effects of two different heuristics on two different possible newscast structures. This model allows us to avoid a long and expensive manual annotation. The evaluation of the output segmentation is performed automatically using a specifically defined metric.