Adversarial Attacked Teacher for Domain Adaptive Object Detection Under Poor Visibility Conditions

Camera-based object detection encounters challenges in adverse weather, which can compromise the robustness of the perception module within autonomous driving systems. Cutting-edge domain adaptive object detection methods use the teacher-student framework and domain adversarial learning to generate domain-invariant pseudo-labels for self-training. However, the pseudo-labels generated by the teacher model often exhibit a bias toward the majority class, incorporating overconfident false positives and underconfident false negatives. We reveal that pseudo-labels vulnerable to adversarial attacks are more likely to be of low quality. To address this issue, we propose a simple yet effective framework named Adversarial Attacked Teacher (AAT) to improve pseudo-label quality. Specifically, we apply adversarial attacks on the teacher model, prompting it to generate adversarial pseudo-labels to correct bias, suppress overconfidence, and encourage underconfident proposals. We introduce an adaptive pseudo-label regularization to emphasize the influence of pseudo-labels with high certainty and reduce the negative impacts of uncertain predictions. Moreover, reliable minority pseudo-labels, verified by pseudo-label regularization, are oversampled to minimize dataset imbalance without introducing false positives. AAT establishes a new state-of-the-art, achieving 53.0 mAP on the Cityscapes to Foggy Cityscapes benchmark. The code is publicly available at https://github.com/KIT-MRT/AAT/.