Exploring global diverse attention via pairwise temporal relation for video summarization

Abstract Video summarization is an effective way to facilitate video searching and browsing. Most of existing systems employ encoder-decoder based recurrent neural networks, which fail explicitly diversify the system-generated summary frames while requiring intensive computations. In this paper, we propose efficient convolutional network architecture for SUMmarization via Global Diverse Attention called SUM-GDA, adapts attention mechanism in a global perspective consider pairwise temporal relations frames. Particularly, GDA module has two advantages: (1) it models within paired as well among all pairs, thus capturing across one video; (2) reflects importance each frame whole video, leading diverse on these Thus, SUM-GDA beneficial generating form satisfactory summary. Extensive experiments three data sets, i.e., SumMe, TVSum, VTW, have demonstrated that its extension outperform other competing state-of-the-art methods with remarkable improvements. addition, proposed can be run parallel significantly less computational costs, helps deployment highly demanding applications.