SumMe
收藏资源简介:
SumMe 数据集是一个视频摘要数据集,由 25 个视频组成,每个视频都至少标注了 15 个人类摘要(总共 390 个)。本文提出了一种新颖的方法和视频摘要的新基准。因此,我们专注于用户视频,它们是包含一组有趣事件的原始视频。我们的方法首先使用针对原始视频量身定制的新颖“超帧”分割来分割视频。然后,我们使用一组低级、中级和高级特征来估计每个超帧的视觉趣味性。基于此评分,我们选择超帧的最佳子集来创建信息丰富且有趣的摘要。引入的基准带有多个人工创建的摘要,这些摘要是在受控心理实验中获得的。这些数据为客观评估摘要方法和获得视频摘要的新见解铺平了道路。在评估我们的方法时,我们发现它产生了高质量的结果,可以与人工创建的摘要相媲美。
The SumMe dataset is a video summarization dataset consisting of 25 videos, each annotated with at least 15 human-generated summaries, totaling 390 in all. This work proposes a novel method and a new benchmark for video summarization. We therefore focus on user-generated videos, which are raw videos containing a set of engaging events. Our method first segments videos using a novel "superframe" segmentation tailored specifically for raw videos. Subsequently, we estimate the visual interestingness of each superframe using a set of low-level, mid-level, and high-level features. Based on these scores, we select the optimal subset of superframes to create informative and engaging summaries. The proposed benchmark includes multiple manually generated summaries obtained from controlled psychological experiments. This benchmark paves the way for the objective evaluation of summarization methods and the acquisition of new insights into video summarization. When evaluating our proposed method, we find that it yields high-quality results that are comparable to manually created summaries.




