PIPAL stands for Perceptual Image Processing ALgorithms, which is an IQA dataset for perceptual image processing algorithms. Due to the rapid development of image reconstruction (IR) algorithms, especially the emergence of some GAN-based models, existing image quality assessment (IQA) methods can no longer evaluate these image reconstruction methods well. Therefore, the team from the Shenzhen Campus of the Chinese University of Hong Kong proposed this dataset, in which the IQA method should evolve and update along with the IR algorithm, and used the Elo scoring system to compare two images and update the scores.

The training set of this dataset includes 200 reference images, 40 distortion types, 23k distortion images, and more than 1 million human ratings.