Quasi-Static Video Set

我们以与中类似的方式准备训练数据集。从本研究中使用的37个高分辨率Vimeo视频中，将它们下采样2倍后，我们提取了40,000大小为I × 256 × 3的随机裁剪序列，其中I ≥ 12。从序列中排除定界关键帧。在训练时，我们对长度为12的这些作物的随机子序列进行了采样。通过排除第一帧和最后一帧，我们获得长度为10的地面真相 (GT) 序列。采样序列的第一帧和最后帧用于在开始时产生x-1，在结束时产生x10。还采用数据增强 (随机翻转/换位)。我们还准备了4个序列的验证集。它们来自对物体运动没有限制的视频，每个视频的数量在30到50帧之间。我们引入了一组新的长序列测试集，其中相机是准静态的并且前景对象会移动。该数据集将补充现有数据集 (Vid4，REDS和Vimeo-90K)，这些数据集仅包含短或呈现快速场景运动的视频。为了生成这个新的数据集，我们m vimeo.com一个d youtube.com下载视频，并提取4个序列，其中包含准静态场景和运动对象。其中前两个分别为Full HD和HD Ready，另外两个为4K。HD和4k序列分别被下采样2和4倍。这4个序列在帧数量上分别具有以下长度: 379、379和172。它们构成了我们称为准静态视频集的测试数据集。我们将序列的长度限制为379，以确保数据集的同质性，但是包含第一个序列的视频包含更大数量的帧。因此，我们还准备了第一个序列的更长版本，称为序列1-XL。后者包含8782帧。

We prepared the training dataset in a manner similar to prior related work. From the 37 high-resolution Vimeo videos used in this study, after downsampling them by a factor of 2, we extracted 40,000 randomly cropped sequences with dimensions of I × 256 × 3, where I ≥ 12. Delimiting key frames were excluded from these sequences. During training, we randomly sampled subsequences of length 12 from these cropped samples. By removing the first and last frames of each 12-length subsequence, we obtained ground truth (GT) sequences of length 10. The first and last frames of the sampled subsequence were used to generate x_{-1} at the start and x_{10} at the end, respectively. Data augmentation techniques including random flipping and transposition were also adopted. We also prepared a validation set comprising 4 sequences. These sequences were sourced from videos with no constraints on object motion, with each video containing between 30 and 50 frames. We introduced a new long-sequence test dataset where the camera is quasi-static and foreground objects move. This dataset will supplement existing datasets such as Vid4, REDS, and Vimeo-90K, which only include videos with short sequences or rapid scene movements. To build this new dataset, we downloaded videos from vimeo.com and youtube.com, and extracted 4 sequences featuring quasi-static scenes and moving objects. The first two sequences are Full HD and HD Ready respectively, while the other two are 4K resolution. The HD and 4K sequences were downsampled by 2x and 4x, respectively. The frame counts of these 4 sequences are 379, 379, and 172 respectively. These sequences form the test dataset we call the Quasi-Static Video Set. We limited the sequence length to 379 to ensure the homogeneity of the dataset, but the video corresponding to the first sequence has a larger number of frames. Therefore, we also prepared a longer version of the first sequence, named Sequence 1-XL. This longer version contains 8,782 frames.

搜集汇总

背景与挑战

背景概述

Quasi-Static Video Set是一个视频超分辨率测试数据集，包含4个准静态场景和运动对象的视频序列，帧数从172到379不等，并提供一个更长的8782帧序列。该数据集补充了现有数据集，专注于长序列和准静态场景的视频数据。