Perceptual Image Processing Algorithms (PIPAL) dataset

PIPAL数据集是由香港中文大学（深圳）数据科学学院创建的一个大规模图像质量评估数据集，专注于感知图像恢复。该数据集包含29,000张图像，其中包括250张高质量参考图像，每张图像有116种不同的失真。PIPAL数据集特别之处在于包含了基于生成对抗网络（GAN）的方法的结果，这是以往数据集中所缺乏的。数据集通过使用更可靠的“Elo系统”收集了超过1.13百万的人类判断来分配主观分数。基于PIPAL数据集，研究者提出了新的基准，用于图像质量评估（IQA）和超分辨率方法，旨在解决现有IQA方法无法公平评估基于GAN的图像恢复算法的问题。

The PIPAL dataset is a large-scale image quality assessment dataset focused on perceptual image restoration, created by the School of Data Science, The Chinese University of Hong Kong, Shenzhen. It contains a total of 29,000 images, including 250 high-quality reference images, with each reference image paired with 116 distinct distorted variants. A key feature of the PIPAL dataset is that it includes results from generative adversarial network (GAN)-based methods, which were absent in previous datasets. Over 1.13 million human judgments were collected using the more reliable Elo system to assign subjective scores. Based on the PIPAL dataset, researchers have proposed new benchmarks for image quality assessment (IQA) and super-resolution methods, aiming to address the issue that existing IQA methods cannot fairly evaluate GAN-based image restoration algorithms.