Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Haochen Li, Rui Zhang, Hantao Yao, Xin Zhang, Yifan Hao, Xinkai Song, Xiaqing Li, Yongwei Zhao, Ling Li, Yunji Chen

TL;DR

Comprehensive experiments over multiple DAOD tasks show that DA-Ada can efficiently infer a domain-aware visual encoder for boosting domain adaptive object detection.

Abstract

Domain adaptive object detection (DAOD) aims to generalize detectors trained on an annotated source domain to an unlabelled target domain. As the visual-language models (VLMs) can provide essential general knowledge on unseen images, freezing the visual encoder and inserting a domain-agnostic adapter can learn domain-invariant knowledge for DAOD. However, the domain-agnostic adapter is inevitably biased to the source domain. It discards some beneficial knowledge discriminative on the unlabelled domain, i.e., domain-specific knowledge of the target domain. To solve the issue, we propose a novel Domain-Aware Adapter (DA-Ada) tailored for the DAOD task. The key point is exploiting domain-specific knowledge between the essential general knowledge and 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 information discarded by the visual encoder. Comprehensive experiments over multiple DAOD tasks show that DA-Ada can efficiently infer a domain-aware visual encoder for boosting domain adaptive object detection. Our code is available at https://github.com/Therock90421/DA-Ada.

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

TL;DR

Comprehensive experiments over multiple DAOD tasks show that DA-Ada can efficiently infer a domain-aware visual encoder for boosting domain adaptive object detection.

Abstract

Domain adaptive object detection (DAOD) aims to generalize detectors trained on an annotated source domain to an unlabelled target domain. As the visual-language models (VLMs) can provide essential general knowledge on unseen images, freezing the visual encoder and inserting a domain-agnostic adapter can learn domain-invariant knowledge for DAOD. However, the domain-agnostic adapter is inevitably biased to the source domain. It discards some beneficial knowledge discriminative on the unlabelled domain, i.e., domain-specific knowledge of the target domain. To solve the issue, we propose a novel Domain-Aware Adapter (DA-Ada) tailored for the DAOD task. The key point is exploiting domain-specific knowledge between the essential general knowledge and 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 information discarded by the visual encoder. Comprehensive experiments over multiple DAOD tasks show that DA-Ada can efficiently infer a domain-aware visual encoder for boosting domain adaptive object detection. Our code is available at https://github.com/Therock90421/DA-Ada.

Paper Structure

This paper contains 31 sections, 16 equations, 6 figures, 18 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Visual-Languague models
  4. Domain Adaptive Object Detection (DAOD)
  5. Tuning method for VLM
  6. Methodology
  7. Overview
  8. Domain-Invariant Adapter (DIA)
  9. Domain-Specific Adapter (DSA)
  10. Visual-guided Textual Adapter (VTA)
  11. Optimization Objective
  12. Experiment
  13. Datasets and Implementation
  14. Comparison to SOTA methods
  15. Ablation Studies
  16. ...and 16 more sections

Figures (6)

  • Figure 1: (a) Traditional DAOD methods optimize the backbone adversarially. (b) Domain-agnostic adapter is inserted into the frozen visual encoder to learn domain-invariant knowledge. (c) Domain-aware adapter can simultaneously capture the domain-specific knowledge from the discarded feature. (d) The mAP($\%$) comparison on the Cross-Weather Adaptation. Compared with original VLM, domain-agnostic adapter brings significant improvement to the source domain but limited improvement to the source domain, while domain-aware adapter brings significant improvement to both source domain and target domain.
  • Figure 2: Overview of the proposed (a) DA-Ada for DAOD and the architecture of (b) the $i$-th domain-aware adapter module (c) the visual-guided textual adapter.
  • Figure 3: Comparison between (a) DA-Pro and (b) Visual-guided textual adapter.
  • Figure 4: Detection comparison on the Cross-Weather adaptation scenario. We visualize the ground truth (a), the detection boxes of SOTA DA-Pro DA-Pro(c) and our methods (b)(d). mAP: mean Average Precision on the example image
  • Figure 5: Feature comparison on the Cross-Weather adaptation scenario. We visualize (a) the target image and the output feature of (b) traditional adapter, (b) domain-invariant adapter (DIA), (c) domain-specific adapter (DSA) and (d) domain-aware adapter (DA-Ada).
  • ...and 1 more figures