Deployment of a GAN architecture to improve the realism of an urban traffic network

In recent years, supervised learning has made significant results in the domain of computer vision, progressively changing our daily life with the help of Convolutional Neural Networks. Comparatively, unsupervised learning has received less attention, yet the invention of Generative Adversarial Network in 2014 provided a new attracting alternative to traditional representation learning. In this master thesis, we will first try to understand the original GAN structure, its improvements with WGAN-GP, and its limits, and successfully propose qualitative generated samples of size (128, 128) for classical data augmentation problem. About its limitations, we show the mode collapse and computation time problems such networks have. We then tackle this work's primary goal, i.e., deploying such architecture to improve the photorealism of images taken in a virtual urban traffic environment. We try to run a specific GAN that modifies the typical architecture to take as input metadata available from the game engine, pixel-level annotations, and the normal synthetic data. Even though we failed to make such a visual shift due to problems with the annotation, we propose and describe several potential solutions for future work, including CycleGAN and its synthetic-to-photorealistic data translation version CyCADA.