Inferring actor communities from videos

In recent years there has been a growing interest in inferring social relations amongst actors in a video using audiovisual features, co-appearance features or both. The discovered relations between actors have been used for identifying leading roles, detecting rival communities in a movie plot etc. In this paper we propose an unsupervised method which uses the video's transcript and closed caption information for discovering actor communities (group of actors or characters in a film that share a common perspective/viewpoint on an issue) from videos. The method proposed groups together actors using a topic model based approach, which jointly models actor-actor interaction (two actors interact when they share the same scene) and the topics associated with their conversations/dialogs. This joint modeling approach shows encouraging results compared to existing methods. Copyright © 2013 ISCA.