Performance Review of Generative Adversarial Network for a Bi-directional Task

Abstract

Generative Adversarial Networks (GAN) contributed many significant works in computer vision tasks in different research areas. But, to author’s knowledge, there is no research discussion about GAN’s performance in a bi-directional task. In this paper, we utilize Pix2pix network as a GAN example to test its performance in a bi-directional task, which is to transfer daylight image to night image and transfer night image back to daylight image. The experimental results review both success cases and fail cases to get several interesting observations regarding the influence of human’s perception in evaluation.