Probabilistic approach to content-based indexing and categorization of temporally aggregated shots in news videos

Abstract

The most frequently stored and browsed videos in Web video collections, TV shows archives, documentary video archives, video-on-demand systems, personal video archives, etc. are broadcast news videos and sports news videos. Content-based indexing of news videos is based on the automatic detection of shots, i.e. of the main structural video units. Video shots can be of different categories such as intro and final animation, chart, diagram or table shots, anchor, reporter, statement, or interview shots, and finally the most informative report shots. The content analysis of a video shot is a very time-consuming process using specific strategy adequate for a given shot category. To analyse faster the content of videos it is desirable to reduce the video space analysed in time-consuming content-based indexing by using temporal aggregation. The temporal aggregation results in grouping of shots of the same event or the same category into scenes. Furthermore, the determination on the basis of time relations of the most likely category also reduces the analysis time enabling us to apply the adequate method of analysis as the first. The paper examines the usefulness of the time relations of shots to determine the most likely category of a shot and to optimize the order of applied strategies.