DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection

  • 2024-10-11 18:20:04
  • Haochen Li, Rui Zhang, Hantao Yao, Xin Zhang, Yifan Hao, Xinkai Song, Xiaqing Li, Yongwei Zhao, Ling Li, Yunji Chen
  • 0

Abstract

Domain adaptive object detection (DAOD) aims to generalize detectors trainedon an annotated source domain to an unlabelled target domain. As thevisual-language models (VLMs) can provide essential general knowledge on unseenimages, freezing the visual encoder and inserting a domain-agnostic adapter canlearn domain-invariant knowledge for DAOD. However, the domain-agnostic adapteris inevitably biased to the source domain. It discards some beneficialknowledge discriminative on the unlabelled domain, i.e., domain-specificknowledge of the target domain. To solve the issue, we propose a novelDomain-Aware Adapter (DA-Ada) tailored for the DAOD task. The key point isexploiting domain-specific knowledge between the essential general knowledgeand domain-invariant knowledge. DA-Ada consists of the Domain-Invariant Adapter(DIA) for learning domain-invariant knowledge and the Domain-Specific Adapter(DSA) for injecting the domain-specific knowledge from the informationdiscarded by the visual encoder. Comprehensive experiments over multiple DAODtasks show that DA-Ada can efficiently infer a domain-aware visual encoder forboosting domain adaptive object detection. Our code is available athttps://github.com/Therock90421/DA-Ada.

 

Quick Read (beta)

loading the full paper ...