netflix/Vera-Layered-Video-Dataset
收藏资源简介:
该数据集名为Vera分层视频数据集,专为Vera分层扩散模型设计,用于内容保持的视频编辑任务。数据集包含训练集和测试集,训练集分为49帧(3秒)和81帧(5秒)两种序列长度,其中49帧训练集总样本数为11,996,包括背景替换和对象添加两种编辑类型,具体划分如:realistic-set1-bg-change(914样本)、realistic-set1-obj-add(470样本)、realistic-set2-obj-add(770样本)、synthetic-bg-change(4,994样本)和synthetic-obj-add(4,848样本);81帧训练集总样本数为6,005,类似地包含背景替换和对象添加类型。测试集总样本数为141,分为bg-change(69样本)和obj-add(72样本)。数据来源于多个公开数据集,训练集包括Pexels、Mixkit和VideoMatte240K,测试集额外增加了DAVIS和VACEBench。该数据集支持视频生成和编辑研究,尤其适用于分层扩散框架,以分离生成内容与保留内容。
The Vera Hierarchical Video Dataset is a specialized dataset designed for the Vera Hierarchical Diffusion Model, targeting content-preserving video editing tasks. The dataset comprises a training set and a test set. The training set is split into two sequence lengths: 49 frames (3 seconds) and 81 frames (5 seconds). The 49-frame training subset contains a total of 11,996 samples, covering two editing categories: background replacement and object addition, with specific subdivisions as follows: realistic-set1-bg-change (914 samples), realistic-set1-obj-add (470 samples), realistic-set2-obj-add (770 samples), synthetic-bg-change (4,994 samples), and synthetic-obj-add (4,848 samples). The 81-frame training subset has a total of 6,005 samples, which also includes the aforementioned two editing categories. The test set consists of 141 samples in total, divided into bg-change (69 samples) and obj-add (72 samples). The dataset is sourced from multiple public datasets: the training set is collected from Pexels, Mixkit, and VideoMatte240K, while the test set additionally incorporates DAVIS and VACEBench. This dataset supports research on video generation and editing, and is particularly well-suited for hierarchical diffusion frameworks that separate generated content from preserved original content.




