KS-Gen
收藏资源简介:
KS-Gen(也称为SkillVid)是一个针对野外人类技能关键步骤生成(KS-Gen)的基准数据集。该数据集旨在支持人类技能视频在关键步骤级别的生成,使研究人员能够开发出能够根据文本描述生成连贯技能演示序列的模型。数据集包含标注文件和视频数据,标注文件以parquet格式提供,包括训练集和测试集的注释,以及来自CrossTask、HT-Step和Kinetics-400数据集的额外注释。每个标注条目包括视频文件路径、关键步骤的起始和结束时间戳、文本描述、帧率、视频字幕、关键帧字幕、大型语言模型生成的字幕、唯一标识符、视频分辨率以及各种评分指标。
KS-Gen (also known as SkillVid) is a benchmark dataset for key step generation of real-world human skills. This dataset aims to support the generation of human skill videos at the key step level, enabling researchers to develop models that can generate coherent skill demonstration sequences based on text descriptions. The dataset includes annotation files and video data. The annotation files are provided in Parquet format, covering annotations for the training and test sets, as well as additional annotations from the CrossTask, HT-Step, and Kinetics-400 datasets. Each annotation entry includes the video file path, start and end timestamps of key steps, text descriptions, frame rate, video subtitles, key frame subtitles, subtitles generated by large language models, unique identifiers, video resolution, and various scoring metrics.




