Abstract : We are investigating the problem of speaker and face identification in broadcast videos. Identification is performed by associating automatically extracted names from overlaid texts with speaker and face clusters. We aimed at exploiting the structure of news videos to solve name/cluster association ambiguities and clustering errors. The proposed approach combines iteratively two conditional random fields (CRF). The first CRF performs the person diarization (joint temporal segmentation, clustering, and association of voices and faces) jointly over the speech segments and the face tracks. It benefits from contextual information being extracted from the image backgrounds and the overlaid texts. The second CRF associates names with person clusters, thanks to co-occurrence statistics. Experiments conducted on a recent and substantial public dataset containing reports and debates demonstrate the interest and complementarity of the different modeling steps and information sources: the use of these elements enables us to obtain better performances in clustering and identification, especially in studio scenes.
