Unsupervised domain-adaptive object detection: An efficient method based on UDA-DETR

Object detection based on deep learning has a wide range of applications in everyday life. However, when a domain gap exists between the training data (source domain) and real-world data (target domain), the performance of many object detectors significantly deteriorates. To address this issue, numerous Unsupervised Domain Adaptation (UDA) detectors attempt to reduce domain discrepancies and align cross-domain features. While these methods have achieved some success, they often align global features in a class-agnostic manner, neglecting the differences in feature distributions across categories within each domain. Our approach emphasizes the importance of both global image features and local instance features across domains, as well as category-specific information. Specifically, we propose an Encoder Feature Alignment (EFA) module, which introduces domain queries to adversarially align encoder-generated features, enabling the detector to extract more domain-invariant features. Additionally, we design an Instance-level Feature Alignment (IFA) module that extracts class-specific central features from the decoder for category-aware cross-domain feature alignment. During each training iteration, local class features progressively converge to global class features, guided by contrastive learning and adversarial loss to achieve Global Feature Alignment (GFA). Our method achieves 46.0% mean accuracy (mAP) in the Weather Adaptation scenario. Compared to the baseline model, a 19.1% mAP gain is achieved (26.9% $\to$ 46.0%). Extensive experimental results show that our proposed model achieves excellent detection performance and strong generalization ability on multiple domain adaptation benchmark datasets.