Person identity clustering in TV show videos

The goal of this work is to cluster all occurrences of TV show characters into different sets with respect to their own identity. A two stage method is presented in this paper. First shot layer processing is performed to gather close-ups with the same central person together and to differentiate far full views from close-up views, hence complicated inter shot facial feature matching is avoided. Then their exact position is localized by face detection and tracking for close-up views. For far full views significant clothing colour recovering and matching is presented. Experiments are presented on a segment from the famous French TV show ‘Le Grand Echiquier’ with satisfactory results.