Video Collection Summarization by Semantic Graph Comparison

As the amount of available digital video content is increasing exponentially, novel ways of storing, accessing and retrieving it are being developped, such as indexing, segmentation or abstraction techniques. Video abstraction can be useful in many ways – from automatic home movie editing to easier (and faster) exploration of a video collection. Video summaries can be a set of carefully selected key frames or the concatenation of video skims [1]. Rather than focusing on a single video, we aim at summarizing several videos from a single collection. A video collection can be seen as a homogeneous set of videos, with the same type of structure, style, duration, etc. Given a video collection, the objective is to generate a short video that is a representative of this collection. The representativeness of a video depends on the application (faster browsing or narrative summary, for instance). It is not reasonable to process each element of a collection independently to produce the final collection summary. As a matter of fact, such an approach would not take redundancy between videos into account and there is no way it can uncover the commonalities between episodes. Therefore, our approach considers a collection as a whole and rely on two steps: video structuring and structure comparison.