Cinematic-based Model for Scene Boundary Detection

Most current video retrieval systems use shot as the basic unit for information organization and access. A shot, however, models only a visually contiguous sequence of video frames with no coherent semantic meanings. On the other hand, viewers tend to “view” video contents in terms of episodes or scenes. In cinematography, scene is the basic story unit that the directors use to compose and convey their ideas. This paper proposes a cinematic-based model to analyze video contents and extract scene boundaries. Using the list of shots as the base, we first locate rough scene boundaries based on visual similarity. Next, we employ the cinematic rules computationally to refine and cluster the rough scenes. We use a combination of camera focal length and similar shot sequence information to derive the final set of scenes. We test our system on two videos of about 44 minutes in duration. The system has been found to be effective.