Talking faces indexing in TV-content

摘要

Our objective is to index talking faces in a TV-Context: build a description of TV-content, in terms of talking people, without any pre-defined dictionary of identities. In TV-content, because of multi-face shots and non-speaking face shots, it is difficult to determine which face is speaking. In this work, a method is proposed which clusters people independently by the audio and by the visual information and combines these clusterings of people (audio and visual) in order to detect sequences of talking faces. The audio indexing system is based on agglomerative clustering with the Bayesian Information Criterion. The visual indexing system is based on costume detection and clustering of color histograms. The combination of both indexes is based on searching for the best match between both clusterings, to obtain a correspondence between the automatic audio labels and the automatic video labels. The talking faces are then determined by the intersection of the segments of the associated audio and video labels. Results of experiments on a TV-Show database show that a high correct detection rate can be achieved by the proposed method.