Application of CycleGAN-based Augmentation for Autonomous Driving at Night

Abstract

Self-driving vehicles contain a number of modules allowing them to autonomously navigate in uncertain environment. The robust, efficient, safe and accurate autonomous navigation are heavily depend on parameters of a perception module. In this paper, we consider perception module as a combination of object detection and road segmentation submodules. As a matter of fact, all of them are based on Deep learning technique. It leads to liability of a big training datasets to provide the accuracy, efficiency and robustness of a perception module for a self-driving car operating in a wide range of scenarios. This paper presents the GAN-based augmentation as a key factor allowing to improve the performances of perception. The provided research shows the comparison between classical augmentation method and CycleGAN-based method. The main focus is made on detection and segmentation problems at nights. The initial training data includes BDD100K dataset and our own one collected in winter time by means of front-view camera of a self-driving car developed in Innopolis University. The obtained results show the improvement of segmentation task in case of application of CycleGAN augmentation. However, the chosen method of GAN-based augmentation has not shown the positive influence on object detection due to appeared visual artifacts.