Character identification is an important task for preparing the metadata for a TV-series, since several applications, such as video summarization [5] , analysis of character interactions [7] , and shot retrieval [3, 8] , require knowing the identities of humans in the scenes. Nevertheless, character identification in TV-series remains a challenging task since the video is usually unconstrained and the human pose varies. Existing works use the names provided in the screenplay, speech

more »

... on [6], and attributes (e.g., gender) [1], to assist person identification. Nevertheless, several prior efforts [9, 10] based on face clustering and tracking are not suitable for consistently identifying the characters in a TV-series due to shot variations and the occlusion of faces. Since the human body is more perceivable even when the face is occluded, person tracking [2] can provide additional advantages for person identification.