InstaVariety

除了发布我们的预处理tfrecords (见下文)，其中包含我们在从视频中学习3D人体动力学时引入的确切的InstaVariety数据外，我们还发布了从Instagram下载的完整数据集。特别是，为了生成InstaVariety，我们为每个1000个帖子抓取了92个面向运动的Instagram标签 (完整标签列表)。Instagram帖子可以是视频或照片，因此过滤掉照片后，我们最终得到的视频比每个标签1000少得多。在92个标签中，我们总共收集了28,272个视频。我们提供了一个文件InstaVariety.json，其中包含每个视频的链接，以及每个视频的相关元数据

In addition to releasing our preprocessed tfrecords (see below), which contain the exact InstaVariety data we introduced for learning 3D human dynamics from videos, we also release the full dataset downloaded from Instagram. Specifically, to generate the InstaVariety dataset, we crawled 92 motion-oriented Instagram tags, with up to 1000 posts per tag (full tag list provided). Instagram posts can be either videos or photos, so after filtering out photo posts, we ended up with far fewer than 1000 videos per tag. Across the 92 tags, we collected a total of 28,272 videos. We provide a file named InstaVariety.json, which contains the links to each video as well as the relevant metadata for each video.