GAN Quality Index (GQI) By GAN-induced Classifier

Yuancheng Ye, Lijuan Wang, Yue Wu, Yinpeng Chen, Yingli Tian, Zicheng Liu, Zhengyou Zhang

Abstract: We propose an objective measure, called GAN Quality Index (GQI), to evaluate GANs. The idea is to train a GAN-induced classifier from the GAN generated data and use its accuracy on a real test set as a metric to measure how well the GAN model distribution matches the real data distribution. Unlike most existing quantitative measurements of GANs, which only reflect partial characteristics of generation distribution, the accuracy of the GAN-induced classifier can be used to derive a simple yet sufficient index to measure how well the generation distribution matches the true data distribution. We demonstrate the effectiveness of GQI on CIFAR-100, Flower-102, and MS-Celeb-1M which contains 10,000 classes.

Keywords: GAN quality index, GAN-induced classifier

TL;DR: We propose an objective measure, called GAN Quality Index (GQI), to evaluate GANs.