Automatic Generation of Personalized Video Summary Based on Context Flow and Distinctive Events

Abstract

This paper proposes an automatic generation method of video summary which can dynamically change the scene component ratio in relation to context semantics and to individual distinctive events. We implemented a tennis digest generation system. In order to convey the context flow of the whole tennis match, two kinds of information were used; one was structural information such as the boundary of each set and game in the match, and the other was each player’s superiority value, which could be calculated at any given time in the match. Also, in order to show individual distinctive scenes such as great moments of play and fantastic shots, indices corresponding to the player’s actions were used. The importance value of the summary candidates was dynamically re-calculated according to the user’s viewpoint and preferences, such as their favorite players and the context scenario of the final summary. Then, the summary candidates were reconfigured to generate the final video and narration text, which were to be played back in the range of the given summary time. Experimental results show that the generated summaries can convey the semantics of the original video reasonably well, and they demonstrate the performance and the validity of our approach.