Cross-Domain Object Detection for Autonomous Driving: A Stepwise Domain Adaptative YOLO Approach

Supervised object detection models based on deep learning technologies cannot perform well in domain shift scenarios where annotated data for training is always insufficient. To this end, domain adaptation technologies for knowledge transfer have emerged to handle the domain shift problems. A stepwise domain adaptive YOLO (S-DAYOLO) framework is developed which constructs an auxiliary domain to bridge the domain gap and uses a new domain adaptive YOLO (DAYOLO) in cross-domain object detection tasks. Different from the previous solutions, the auxiliary domain is composed of original source images and synthetic images that are translated from source images to the similar ones in the target domain. DAYOLO based on YOLOv5s is designed with a category-consistent regularization module and adaptation modules for image-level and instance-level features to generate domain invariant representations. Our proposed method is trained and evaluated by using five public driving datasets including Cityscapes, Foggy Cityscapes, BDD100K, KITTI, and KAIST. Experiment results demonstrate that object detection performance is significantly improved when using our proposed method in various domain shift scenarios for autonomous driving applications.