AI to generate photorealistic simulated sensor data in physically accurate simulated environments

To be useful and efficient, simulation environments are required to be highly realistic. Namely, it is necessary to reduce the gap between simulated and real-world environments and to obtain simulation sensor output data close to realistic sensor ones. However, obtaining highly realistic simulation environments requires a massive computational and resource effort. An efficient alternative strategy to overcome this issue, consists in exploiting AI to enhance the realism of simulated sensor data. Indeed, the recent advances of AI methods, in the last years, provide innovative models that can be used to address the issue of enhancing the realism of simulation data. Specifically, Generative Adversarial Networks (GANs) are an AI model widely employed in this field. The research activity presented in the thesis starts from a recent proposed AI framework, that is introduced to translate synthetic images data into photorealistic urban scenes. The research field that considers the problem of translating images from one domain to another one is addressed as image-to-image translation, while its video counterpart is addressed as video-to-video synthesis. In the thesis work, the state-of-the-art method is practically implemented to reach resulting enhanced data close to the ones presented by the authors. Subsequently, different implementations of the framework are proposed, to work with different synthetic data generated with a 3D physically accurate simulator. Finally, the results, obtained implementing the framework, are qualitatively evaluated and compared.