An Alternative Lightness Control with GAN for Augmenting Camera Data

To build an autonomous car, many technologies have to be taken into application. The most important component of a fully self-driving car is the object detection system. This system is responsible for detecting obstacles on the street. However, these detection models still face many difficulties such as unable to work on extreme conditions (storm, night, chaotic road,…). To tackle one aspect of this problem, in this paper we propose an augmentation method that creates more data by generating night images from day images and vice versa using LCcycleGAN, a Lightness conditional Unpaired Image-to-Image Translation approach, this framework is the fusion of CycleGAN [1] and conditional GAN [2]. To evaluate our method, we measure performance of YoloV3 [3] on our collected dataset (and augmented data) consists of day and night images of Vietnamese streets which are often highly chaotic and extreme. Our method increases AP of base vehicle detection model's performance from 0.5 to 0.56.