QST (Quick Sky Time)

QST包含从YouTube收集的216延时4k视频中切出的1,167视频片段，可用于多种任务，例如 (高分辨率) 视频生成，(高分辨率) 视频预测，(高分辨率) 图像生成，纹理生成，图像修复、图像/视频超分辨率、图像/视频着色、图像/视频动画等。每个短片包含多个帧 (从最小58帧到最大1,200帧，总共285,446帧)，每个帧的分辨率超过1,024x1,024。具体地，QST由训练集 (包含1000个剪辑，总共244,930个帧) 、验证集 (包含100个剪辑，总共23,200个帧) 和测试集 (包含67个剪辑，总共17,316个帧) 组成。

QST contains 1,167 video clips extracted from 216 4K time-lapse videos collected from YouTube. It supports a wide range of downstream tasks, including (high-resolution) video generation, (high-resolution) video prediction, (high-resolution) image generation, texture generation, image inpainting, image/video super-resolution, image/video colorization, image/video animation, etc. Each short clip consists of multiple frames, with the frame count ranging from a minimum of 58 to a maximum of 1,200, and the total number of frames across all clips reaches 285,446. The resolution of each frame exceeds 1,024×1,024. Specifically, QST is split into three subsets: a training set with 1,000 clips and a total of 244,930 frames, a validation set with 100 clips and a total of 23,200 frames, and a test set with 67 clips and a total of 17,316 frames.

搜集汇总

背景与挑战

背景概述

QST数据集包含1,167个从216个4K延时视频中提取的片段，总计285,446帧，分辨率超过1,024x1,024，适用于视频生成、预测、图像修复等多种任务。数据集分为训练集、验证集和测试集，由浙江大学和湖州大学于2020年发布。