Unsupervised domain adaptive object detection via discriminative instance teacher

Domain adaptive object detection (DAOD) poses significant challenges due to pronounced domain shifts. Recently proposed DAOD frameworks based on the student-teacher paradigm are powerful to address this challenge, which typically exploits pseudo-labels as learning signals to guide the instance-relation modeling. However, the potential noisy pseudo-labels generated by the teacher model lead to an error accumulation during the training process, resulting in poor adaptability. Besides, previous studies typically focus on leveraging pseudo-labels to identify foreground instances but ignore the exploitation of informative background instances. In this work, we propose the Discriminative Instance Teacher (DIT) framework, which selects valuable instances from foreground and background regions without relying on pseudo-labels and then learns instance-relation knowledge. Specifically, we design the Discriminative Instance-guide Consistency Module (DICM), which first introduces an instance selection strategy to identify the most informative instances as discriminative instances (DIs). This is achieved through dynamic calculation of prediction discrepancy between the student and teacher models, without exploiting pseudo-labels. Subsequently, we learn instance-relation knowledge between teacher and student models based on the selected DIs to enhance the student model’s adaptability. Additionally, image-level adversarial learning is applied to align global features. Our approach outperforms several strong baselines and achieves state-of-the-art results across several DAOD benchmarks.