This paper identifies shortcomings of existing benchmarking approaches in the domain-adaptive object detection (DAOD) field and presents a novel benchmarking framework, training and evaluation protocols, benchmark datasets, and an algorithm called ALDI++ that achieves new state-of-the-art performance to address them. The shortcomings of existing studies include underestimation of baseline performance, inconsistent implementations, outdated backbone networks, and lack of diversity in benchmark datasets. The proposed ALDI++ significantly outperforms existing state-of-the-art approaches in various domain adaptation scenarios, such as Cityscapes to Foggy Cityscapes, Sim10k to Cityscapes, and CFC Kenai to Channel. The ALDI framework can also be applied to YOLO and DETR-based DAOD, and achieves state-of-the-art performance without additional hyperparameter tuning. This study redefines benchmarking in the DAOD field and lays the foundation for future research.
