Cross-domain Object Detection with Dual-path Consistency Constraints

In practical scenarios such as intelligent traffic surveillance, object detection models often suffer from performance degradation when training data and deployment environments exhibit significant domain shifts caused by differences in imaging devices, scene conditions, and the weather. To improve the stability of unsupervised domain adaptive object detection, in this research, a cross-domain object detection method termed regularized mean teacher (RMT) is proposed based on dual-path consistency constraints. Built upon the mean teacher framework, RMT introduces a semantic-level consistency branch and an instance-level consistency branch during training. The semantic branch constrains multi-scale feature representations between the teacher and student models, while the instance branch establishes consistency constraints on matched object-level features. These two branches provide auxiliary supervision beyond pseudo labels and reduce the influence of unstable teacher predictions under domain shift. The proposed branches are used only during training, and no additional inference branch or forward computation is introduced during testing. Experiments on the Cityscapes to Foggy Cityscapes benchmark show that RMT achieves 53.1% mAP, outperforming the baseline model and several representative methods. The results demonstrate that dual-path consistency constraints can enhance teacher-student representation stability and improve detection performance under adverse weather conditions.