Property Constrained Video Summarization via Regret Minimization

Video summarization has become one of the most effective solutions for quickly understanding a large amount of video data. Video properties such as importance, diversity, representativeness, and storyness have been widely adopted for summarization based on kinds of features of video frames. To fully exploit these properties, in this paper, we propose a property constrained video summarization framework to output fixed-size summaries based on the concept of regret minimization which is popular in the database community for solving multi-criteria decision-making problems. Our framework first uses representativeness property and effective techniques to generate a candidate frame set. Then, the candidate set is transformed into a multi-dimensional point set where dimensions are composed of entities of the importance property and the diversity property. Moreover, our framework runs the regret minimization query on the point set to output a summary. To further improve the quality of the output summary, the property storyness is exploited to add frames satisfying the size constraint. Experimental results on the benchmark datasets demonstrate the effectiveness and superiority of our framework.