History-aware adaptive teacher for cross-domain object detection

The teacher-student self-training framework has shown promising results in cross-domain object detection, which utilizes pseudo-labels generated by the teacher model to guide the student model in learning target-domain knowledge. However, existing teacher-student frameworks lack a safeguard against irreversible collapse caused by pseudo-label noise and domain shift, making the system vulnerable to catastrophic degradation. To address this issue, we propose History-aware Adaptive Teacher (HAT), a novel self-training framework that integrates historical knowledge to stabilize teacher model updates and generate reliable pseudo-labels, enabling robust domain adaptation. Specifically, we design a Dynamic History-aware Teacher Stabilization (DHTS) module, which preserves historical optimal model states and uses them as adaptive references to guide teacher updates, preventing irreversible collapse. Building upon this, we devise a Dual-Consensus Pseudo-label Refinement (DCPR) module to optimize the pseudo-label generation process through cross-teacher localization stability and category consistency. Furthermore, we incorporate a Multi-Scale Adversarial Feature Alignment (MSAFA) module into the FCOS detector, enabling layer-wise adversarial learning for precise feature distribution alignment. Extensive experiments conducted on three standard benchmarks demonstrate the superiority of our method. Most notably, our method achieves 42.8% mAP on the challenging Cityscapes to BDD100K-Daytime adaptation scenario, marking a substantial improvement of 9.5% mAP over state-of-the-art methods. The source code will be made publicly available upon acceptance.