Reliability of GAN Generated Data to Train and Validate Perception Systems for Autonomous Vehicles

Abstract

Autonomous systems deployed in the real world have to deal with potential problem causing situations that they have never seen during their training phases. Due to the long-tail nature of events, collecting a large amount of data for such corner cases is a difficult task. While simulation is one plausible solution, recent developments in the field of Generative Adversarial Networks (GANs) make them a promising tool to generate and augment realistic data without exhibiting a domain shift from actual real data. In this manuscript, we empirically analyze and propose novel solutions for the trust that we can place on GAN generated data for training and validation of vision-based perception modules like object detection and scenario classification.