Evaluation of Image Synthesis for Automotive Purposes

摘要

The aim of this article is to evaluate a state of the art image synthesis carried out via Generative Adversarial Networks (conditional Wasserstein GAN and Self Attention GAN) on a traffic signs dataset. For the experiment, we focused on generating images with a 64 x 64-pixel resolution as well as on the GAN's ability to capture structural and geometric patterns. Four different GAN architectures were trained in order to highlight the difficulties of the training, such as collapse mode, vanishing gradient and resulting image fidelity. The Frechent Inception Distance is compared with other state of the art results. The importance of evaluating on automotive datasets as well as additional wishes for further improvements are addressed at the end of this article.