OPFSumm: on the video summarization using Optimum-Path Forest

Abstract

Video summarization attempts at encoding a given video into a compact representation for a better storage and retrieval purposes. This work copes with the problem of static video summarization using the unsupervised Optimum-Path Forest (OPF). We sampled the encoded video sequence into frames and extracted features based on color information or spectral properties. After that, meaningless frames are removed, and OPF models the problem of video summarization as a clustering process. Possible redundant keyframes are filtered, and at last the video summary is created based on non-redundant keyframes. We presented a more in-depth study that also considers temporal information to obtain better video representations. The experiments over three public datasets were analyzed through F-measure evaluation metric and showed the robustness of OPF for automatic video summarization: 0.19 for SumMe dataset, 0.728 concerning Open Video dataset, and 0.451 regarding YouTube dataset..