Enhancing Nighttime Vehicle Detection via Transformer-based Data Augmentation

In autonomous driving systems, vehicle detection technology typically relies on object detection models trained on driving image datasets. However, accurate vehicle detection becomes challenging during nighttime due to low-light conditions, necessitating a sufficient amount of nighttime driving images for training the model. Unfortunately, publicly available datasets lack an adequate amount of nighttime driving images, and collecting them directly is cost-ineffective. In this paper, we propose a novel augmentation method based on transformer to convert daytime driving images into realistic nighttime driving images. Our method analyzes the style case of the given daytime driving image, selects a tailored style image that corresponds to the analyzed style case, and transfers the daytime driving image into the realistic nighttime driving image using the selected style image. Our diverse range of evaluations demonstrates the effectiveness of our proposed method in augmenting realistic nighttime driving images.