- Abstract: We tackle the problem of domain adaptation in object detection, where the main challenge lies in significant domain shifts between source (one domain with supervision) and target (a domain of interest without supervision). Although the teacher-student framework (a student model learns from pseudo labels generated from a teacher model) has been adopted to enable domain adaptation and yielded accuracy gains on the target domain, the teacher model still generates a large number of low-quality pseudo labels (e.g.,false positives) due to its bias toward source domain. This leads to sub-optimal domain adaptation performance. To ad-dress this issue, we propose Adaptive Unbiased Teacher (AUT), a teacher-student framework leveraging adversarial learning (on features derived from backbone)and weak-strong data augmentation to address domain shifts. Specifically, we em-ploy feature-level adversarial training, ensuring features extracted from the source and target domains share similar statistics. This enables the student model to capture domain-invariant features. Furthermore, we apply weak-strong augmentation and mutual learning of the teacher for target domain and student model for both domains. This enables the updated teacher model to gradually benefit from the student model without suffering domain shift. We show that AUT demonstrates superiority over all existing approaches and even Oracle (fully-supervised) mod-els by a huge margin. For example, we achieve 50.9% (49.3%) mAP on FoggyCityscape (Clipart1K), which is 9.2% (5.2%) and 8.2% (11.0%) higher than previous state of the arts and Oracle, respectively.