イベント-アクション・モデルに基づく動画像検索システム: STRIKE

This paper describes a movie retrieval system that enables users to retrieve scenes according to various kinds of viewpoints. In most of traditional movie retrieval systems, contents of scene are modeled as sets of attributes (such as keywords), and user retrieves scenes using these attributes. In this approach, it is hard for users to retrieve scenes according to view of users, because these attributes are prepared based on viewpoints of database manager. We developed event-action model that enables to compose scenes dynamically according to viewpoints of user. In this model, a difference between two continuous frames in a movie is represented as an event, and a context of movie is represented as a sequence of events. For retrievals, users specifies sub-sequence of the contexts, and the system extracts specified sub-sequences and converts them to scenes.