A Framework for Evaluating Video Summary Approaches

Video summarization is a crucial task to solve the explosion of video data. The goal of video summarizing is to create a shortened version of the original video while retaining its essential and pertinent content. In general, a video summary system is composed of three primary modules: shot boundary detection, shot scoring, and shot selection. However, existing research focuses exclusively on a single module, necessitating a comprehensive assessment when methods are changed across modules. In this study, we provide a framework for evaluating alternative techniques for the video summary problem that permits multiple combinations in different modules to evaluate the significance of adding method stages in video summaries. The analysis and combination results of the framework reveal that the combination of Uniform and DSNet Anchor-free provides state-of-the-art performance on the SumMe dataset. We also provide the framework source code ¹ for the community.