Scholarlink

<p>Domain Adaptive Object Detection (DAOD) transfers knowledge from a labeled source domain to an unlabeled target domain via feature alignment. While most methods adopt the Mean-Teacher framework, recent approaches leverage large Visual Foundation Models (VFMs) like DINO-Teacher for pseudo-label generation. However, their alignment performance heavily depends on pseudo-label quality, for which a principled evaluation standard is lacking. Moreover, these methods often fail to distinguish features of easily confused categories, limiting training effectiveness. To address these issues, we propose a large model framework that enhances DAOD through pseudo-label assessment and class-aware feature alignment (PAFA). We introduce a collaborative large-model-based criterion to evaluate and filter pseudo-labels, and propose a Class-Aware Matrix (CAM) to enable fine-grained, category-wise alignment. Our method achieves state-of-the-art results on multiple benchmarks, improving by +1.8%mAP on Foggy Cityscapes and +1.6%mAP on BDD100K Daytime, with a notable +3.6%mAP gain on BDD100K Nighttime, demonstrating strong robustness and generalizability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

PAFA: adapting large models for domain adaptive object detection

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Zixiao Cui,
  • Jie Wu,
  • Mingbin Li,
  • Fan Zhou,
  • Zhuo Su

摘要

Domain Adaptive Object Detection (DAOD) transfers knowledge from a labeled source domain to an unlabeled target domain via feature alignment. While most methods adopt the Mean-Teacher framework, recent approaches leverage large Visual Foundation Models (VFMs) like DINO-Teacher for pseudo-label generation. However, their alignment performance heavily depends on pseudo-label quality, for which a principled evaluation standard is lacking. Moreover, these methods often fail to distinguish features of easily confused categories, limiting training effectiveness. To address these issues, we propose a large model framework that enhances DAOD through pseudo-label assessment and class-aware feature alignment (PAFA). We introduce a collaborative large-model-based criterion to evaluate and filter pseudo-labels, and propose a Class-Aware Matrix (CAM) to enable fine-grained, category-wise alignment. Our method achieves state-of-the-art results on multiple benchmarks, improving by +1.8%mAP on Foggy Cityscapes and +1.6%mAP on BDD100K Daytime, with a notable +3.6%mAP gain on BDD100K Nighttime, demonstrating strong robustness and generalizability.