CSG-405
收藏资源简介:
CSG-405数据集是一个包含405小时高分辨率视频的大型数据集,涵盖了71种常见的演讲类型,并标注有2D骨架和多样化的演讲者人口统计信息。该数据集旨在促进高保真度伴随手势视频生成的研究,其中视频片段经过多阶段过滤,包括骨架检测质量和音频唇同步评估。数据集包含不同性别、年龄和种族的演讲者,覆盖了正式和非正式的演讲场景,并具有丰富的2D骨架注释和高质量的图像分辨率。
The CSG-405 dataset is a large-scale resource containing 405 hours of high-resolution videos, covering 71 common speech types, and annotated with 2D skeletons and diverse speaker demographic information. This dataset aims to advance research on high-fidelity gesture-aware video generation, where video clips undergo multi-stage filtering including skeleton detection quality assessment and audio-visual lip synchronization evaluation. The dataset features speakers of varying genders, ages and ethnicities, spans both formal and informal speech scenarios, and provides rich 2D skeleton annotations along with high-quality video resolution.




