Conditional Drive Environment Translation using StarGAN with CBIN

Automatic driving without human's control requires a lot of training to be able to adopt any situation. About situation, the environment at driving has the variety such as dark at night, wet road because of rain and the stack of snow. Preparing the drive image dataset with a lot of environment situations is the difficult task in collecting. This study tried solving the preparing problem by constructing deep learning model with limited data and translating single image to multi environment situations. For accomplishing this subject, we employed N domains translation model called StarGAN. In this paper, we investigated the StarGAN with enough conditional translation performance through comparing visualized results and FID score. We trained StarGANs including original to translate single image to 6 kinds of drive environment domain: daytime & clear, daytime & rainy, daytime & snowy, night & clear, night & rainy and night & snowy. Through experiments, we found the StarGAN employing “CBIN: Central Biasing Instance Normalization” and “AdaLIN: Adaptive Layer-Instance Normalization” at Generator, and “the adversarial loss of CAM Logit” at Discriminator could mark the lower FID score than original StarGAN.