Adversarial Attacked Teacher for Unsupervised Domain Adaptive Object Detection

TL;DR

This work tackles domain shifts in unsupervised domain adaptive object detection by introducing Adversarial Attacked Teacher (AAT), a mean-teacher framework that uses adversarial perturbations on the teacher to generate adversarial pseudo-labels as complementary supervision. The core ideas—Adaptive Pseudo-label Regularization (APR) and Robust Minority Oversampling (RMO)—regularize learning by emphasizing high-certainty pseudo-labels, correcting bias toward majority classes, and oversampling robust minority objects to balance datasets. Empirical results on Cityscapes→Foggy Cityscapes and PASCAL VOC→Clipart1k show substantial gains, achieving up to 53.0 mAP on Foggy Cityscapes and 52.6 mAP on Clipart1k, surpassing state-of-the-art methods and demonstrating improved performance on minority classes. The method provides a practical, annotation-free improvement to DAOD with wide applicability to adverse-weather and artistic domain shifts.

Abstract

Object detectors encounter challenges in handling domain shifts. 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 tend to be biased towards the majority class and often mistakenly include overconfident false positives and underconfident false negatives. We reveal that pseudo-labels vulnerable to adversarial attacks are more likely to be low-quality. To address this, we propose a simple yet effective framework named Adversarial Attacked Teacher (AAT) to improve the quality of pseudo-labels. Specifically, we apply adversarial attacks to the teacher model, prompting it to generate adversarial pseudo-labels to correct bias, suppress overconfidence, and encourage underconfident proposals. An adaptive pseudo-label regularization is introduced to emphasize the influence of pseudo-labels with high certainty and reduce the negative impacts of uncertain predictions. Moreover, robust minority objects verified by pseudo-label regularization are oversampled to minimize dataset imbalance without introducing false positives. Extensive experiments conducted on various datasets demonstrate that AAT achieves superior performance, reaching 52.6 mAP on Clipart1k, surpassing the previous state-of-the-art by 6.7%.

Figures (6)

• Figure 1: Ground truth (top), vanilla pseudo-labels (middle), and adversarial pseudo-labels (bottom) generated on Foggy Cityscapes. Different colors represent different categories. Vanilla pseudo-labeling tends to exhibit the following issues: (a) bias towards major classes, (b) overconfidence about parts of objects, and (c) underconfidence.
• Figure 2: (a) Overview of the proposed Adversarial Attacked Teacher framework. Our model incorporates a Mean Teacher framework and two key components: Adaptive Pseudo-label Regularization (APR) and Robust Minority Oversampling (RMO). APR involves conducting adversarial attacks on the teacher model to generate adversarial pseudo-labels, which serve as complementary labels by retaining reliable vanilla pseudo-labels and modifying uncertain ones. The student model is trained to fit both types of pseudo-labels. RMO involves oversampling high-certainty objects from minority categories to address class imbalance. (b) Robust Minority Oversampling (RMO) demonstrated on a train object.
• Figure 3: Illustrative examples of pseudo-labels on the original data (blue) and the corresponding adversarial examples (orange).
• Figure 4: Learning curves on Clipart1k dataset.
• Figure 5: Qualitative results on Foggy Cityscapes. We compare AT (top) and AAT (bottom). AAT fixes (a) bias towards major classes, (b) false positives, and (c) false negatives.