Video Highlights

视频亮点数据集包含超过1470个完整的综艺节目视频 (约1200h) 播放链接和包含提取功能的文件。其中，Highlight Clips包括18,000个视频，总时长约为750小时，并且每个视频都被精确标记，Time-stamped.Researchers可以使用训练数据集中的视频和标记的剪辑来设计算法和训练模型。

数据集分为三个子集:

训练集: 可用于设计算法和训练模型。它包含视频功能和标签信息。

验证集: 包含相对较少的视频，可用于验证训练好的模型。此外，还提供了用于突出显示剪辑时间戳的ground-truth (Meta.json)。
测试集: 包含另一组视频 (不带视频亮点时间戳地面真相)。提交预测时间戳后，可用于评估模型性能和排名。

The Video Highlight Dataset provides playback links for over 1,470 full-length variety show videos (approximately 1,200 hours in total) along with files that support extraction functionality. The Highlight Clips subset comprises 18,000 videos with a total duration of roughly 750 hours, each of which is accurately timestamped. Researchers can utilize the videos and labeled clips from the training dataset to design algorithms and train models.

The dataset is divided into three subsets:

1. Training Set: Intended for algorithm development and model training, this subset contains video features and label information.
2. Validation Set: Includes a relatively small number of videos for validating trained models. Additionally, the ground-truth data for highlight clip timestamps is provided in Meta.json.
3. Test Set: Contains another set of videos without the ground-truth of video highlight timestamps. After submitting predicted timestamps, this subset can be used to evaluate model performance and conduct ranking evaluations.

搜集汇总

背景与挑战

背景概述

Video Highlights数据集包含超过1470个完整综艺节目视频（约1200小时）和18,000个带精确时间戳的高亮剪辑（约750小时），用于视频亮点检测。该数据集分为训练集、验证集和测试集，支持算法设计、模型训练和性能评估。