YOLO-D: A Domain Adaptive approach towards low light object detection

Low light object detection in dark environments is very complicated due to the fact that there is always low visibility, low contrast and a lot of noise, which impact on the conventional object detection systems negatively. This paper proposes a Domain Adaptive framework YOLO-D integrating low-light enhancement techniques with multiscale domain adaption in order to improve object detection in difficult illumination conditions. The proposed framework builds on the YOLOv8 architecture by including a Low Light Enhancement Module(LLEM) that preprocesses images to increase exposure and increase their contrast while reducing any noise within them. In order to ensure robustness against domain shifts due to different illumination levels along with a Multiscale Domain Adaptive Network(MS-DAN) is utilized to perform domain adaption at various levels of feature extraction. A strategy for adversarial learning combined with unsupervised learning is used to train the YOLO-D framework so that detection accuracy and domain adaptation are optimized. From experiments done on benchmark dataset ExDark, YOLO-D outperforms other SOTA models, especially under very low light conditions with 97.6% of Mean Average Precision.