An open source quantitative evaluation framework for automatic video summarization algorithms

Abstract

The creation, consumption, and manipulation of video play a central role in everyday life as the amount of video data is growing at an exponential rate. Video summarization consists on producing a condensed output from a video that allows humans to rapidly understand and browse the content of the original source. Although there are several evaluation approaches proposed in the literature, multiple challenges make the quantitative evaluation of a summarization a complex process. In this paper we present a completely open video summarization evaluation framework that is compatible with existing datasets and published results. Standard metrics are considered and a new metric that captures unbalanced-class video summarization evaluation is proposed. Two legacy datasets are integrated in a standard format. Finally, new quantitative results based on already published algorithms are presented.