GAN-Based Unsupervised Day to Night Image Translation for Nighttime Vehicle Detection

A reliable large amount of data is the foundation of deep learning-based target detection tasks. However, the current lack of road datasets in nighttime environments has been a significant hindrance to the development of nighttime vehicle detection. The emergence of adversarial generative networks has made it a promising research direction to translate the existing large amount of daytime images into nighttime images. However, the current translation models are either demanding of the existing data, requiring instance segmentation-level labels; or focused on the general domain. As a result, applying them directly to the day-to-night translation task will produce many unreasonable light spots in the generated images due to nighttime lights. In light of these considerations, this paper proposes a GAN-based unsupervised day-to-night image conversion model, UD2NGAN, which can translate daytime road images to nighttime images without needing segmentation labels. Furthermore, this approach mitigates the occurrence of anomalous light spots in the generated images. The contributions of this paper are threefold. First, we propose a highlight suppression module, which is used to preprocess nighttime images and reduce the number of unreasonable light spots in the generated images. Second, we employ an edge detection module to extract edge features from both the original and the generated images. This results in the retention of more information about the road targets in the generated images. Third, we demonstrate the superiority of our approach by applying the same detectors to multiple datasets.