Shot Selection Strategies in Video News Story Tracking

Abstract: When tracking a news story a user typically chooses a story as input for a query. The computer system then tries to find stories similar to the one used as input. Earlier attempts at video news tracking used all shots in a story, regardless of whether they improved the query or not. Some of those shots are bound to be disruptive to the matching process. Due to this problem with the current methods, we propose a shot selection method which utilizes the similarity between shots within a story to automatically extract shots that are highly representative of the story content. Specifically, three methods are proposed: a) similar shot exclusion; b) similar shot inclusion; and c) length based shot selection. This paper describes experiments performed on Japanese broadcast news video. We find that being restrictive with what shots to include can improve tracking performance. In certain cases, drastically.