4564356346
收藏资源简介:
该数据集是一个用于文本到视频生成任务的多模态数据集,包含超过35,000个同步的音频-视频剪辑。每个数据样本由视频文件(.mp4格式)、对应的音频文件(.m4a格式)和包含元数据的JSON文件构成,以WebDataset的tar包格式组织以便高效流式加载。关键特征是每个剪辑具有足够的时序长度,其潜在表示或帧序列长度(通过公式L=(round(dur*16)-1)//4+1计算)被约束为大于等于15,这有助于模型学习更长的时序依赖关系。数据集适用于训练和评估结合音频与视频信息的生成模型,特别是文本到视频的合成任务。
This dataset is a multimodal dataset for text-to-video generation tasks, containing over 35,000 synchronized audio-visual clips. Each data sample comprises a video file in .mp4 format, a corresponding audio file in .m4a format, and a JSON file containing metadata, and is organized in WebDataset tarball format for efficient streaming loading. The key feature is that each clip has sufficient temporal length, with its latent representation or frame sequence length calculated via the formula L=(round(dur*16)-1)//4+1 constrained to be no less than 15, which helps models learn longer temporal dependencies. This dataset is suitable for training and evaluating generative models that integrate audio and video information, particularly text-to-video synthesis tasks.
数据集概述
- 数据集名称:>=15-latent audio+video clips
- 许可证:未知(license: unknown)
- 任务类别:文本到视频(text-to-video)
数据集详情
- 数据组成:包含同时具有MP4视频和M4A音频的文件,以及对应的JSON文件。
- 筛选条件:
L=(round(dur*16)-1)//4+1 >= 15,该条件用于筛选时长满足特定潜变量长度要求的片段。 - 总片段数:35,028个。
- 存储格式:WebDataset TAR文件,其中每个条目包含三个文件:
{id}.mp4:视频文件{id}.m4a:音频文件{id}.json:元数据文件




