Unsupervised anchorpersons differentiation in news video

Abstract

The automatic extraction of video structure from content is of key importance to enable a variety of multimedia services that span from search and retrieval to content manipulation. An unsupervised independent unimodal clustering method for anchorpersons detection and differentiation in newscasts is presented in this paper. The algorithm exploits audio, frame and face information to identify major cast in the content. These three components are first processed independently during the cluster analysis and then jointly in a compositional mining phase. A differentiation of the role played by the people in the video has been implemented exploiting the temporal characteristics of the detected anchorpersons. Experiments show significant precision/recall results thus opening further research directions in video analysis, particularly when the content is highly structured as in TV newscasts.