Cross-Domain Detection Transformer with Multi-view Adaptive Feature Alignment in Remote Sensing Imagery

Unsupervised Domain Adaptation (UDA) techniques are crucial for remote sensing object detection, designed to address performance degradation caused by the domain gap between training and test data. These methods leverage unlabeled target domain data, thus alleviating the high costs associated with data annotation. Recent developments in Detection Transformers (DETR) have simplified the detection pipeline and attracted significant research interest. Building on this architecture, we introduce an unsupervised domain adaptation detector for remote sensing object detection. Specifically, we introduce a multi-view adaptive feature alignment module that initially captures domain-specific features in complex backgrounds by leveraging a cross-attention mechanism. Subsequently, we employ contrastive learning to enforce the aggregation of domain-specific features from various perspectives, thereby improving the accuracy of feature alignment. Moreover, we demonstrate that integrating the self-training framework into DETR-based detectors can significantly mitigate the domain gap by further utilizing unlabeled data in the target domain. We validated the effectiveness and generalizability of our method across two remote sensing cross-domain detection scenarios using four public datasets.