A relation-enhanced mean-teacher framework for source-free domain adaptation of object detection

Source-Free Domain Adaptation Object Detection (SF-DAOD) is a challenging task in the field of computer vision. This task is used when the source-domain dataset is not accessible. In existing work, three serious issues are not solved: (1) Information on the semantic topological structure among instances is overlooked. (2) In the training process, attention is focused solely on a single domain, without considering the interaction of information between domains. (3) Low-quality pseudo-labels can degrade the training effectiveness. In this paper, we propose a Relation-Enhanced Mean-Teacher (RMT) Framework utilizing graph neural networks to address these issues. We build the graph structure using the semantic topological structure and the location information, and we employ a Graph-Guided Feature Fusion (GFF) network to achieve alignment between the source and target domains. Furthermore, we utilize these features and the graph to construct a Graph-Guide Bidirectional Verification (GBV) to select high-quality pseudo-labels for supervision. Our experiments on four domain shift scenarios with six standard benchmark datasets demonstrate that our approach outperforms various existing state-of-the-art domain adaptation methods.