Fréchet AutoEncoder Distance: A new approach for evaluation of Generative Adversarial Networks

Abstract

Evaluation measures of Generative Adversarial Networks (GANs) have been an active area of research and, currently, there are several measures to evaluate them. The most used GANs evaluation measure is the Fréchet Inception Distance (FID). Measures such as FID are known as model-agnostic methods, where the generator is used as a black box to sample the generated images. Like other measures of model-agnostic, FID uses a deep supervised model for mapping real and generated samples to a feature space. We proposed an approach here with a deep unsupervised model, the Vector Quantised-Variational Autoencoder (VQ-VAE), for estimating the mean and the covariance matrix of the Fréchet Distance and named it Fréchet AutoEncoder Distance (FAED). Our experimental results highlighted that the feature space of the VQ-VAE describes a clustering domain-specific representation more intuitive and visually plausible than the Inception network used by the benchmark FID.