Multi-level consistency regularization for domain adaptive object detection

To improve the adaptability of detectors, most existing domain adaptation algorithms adopt adversarial learning to align feature distributions between source and target datasets. Different from previous methods, this work explores the possibility of transferring detectors with only source domain data and style information of the target domain. Specifically, we propose three consistency regularizations to enhance the adaptation performance of the detector. First, the source domain and the synthetic domain share the same image content, and the supervision regularization fully exploits the source annotations, which narrows the domain gap and saves labeling costs. Second, prediction regularization improves the robustness of the detector to category semantics and location awareness in different domains. Third, self-discovering feature regularization projects the detector’s attention to object-related regions, which are more discriminative than background noise. In addition, our method can cooperate with the classic domain adaptation algorithm to further improve the generalization of the detector, which shows that both the content and style information of target domain images are crucial for the transfer process. Extensive experiments have been conducted on multiple detection benchmarks, including Foggy Cityscapes, Sim10k, KITTI, Clipart, and Watercolor datasets. The favorable performance compared with existing state-of-the-art methods confirms the effectiveness of the proposed consistency regularizations.