[NeurIPS 2025] Domain-RAG: Retrieval-Guided Compositional Image Generation for Cross-Domain Few-Shot Object Detection

📚 Paper (NeurIPS 2025) | 📂 Project Page | 📦 Dataset Scripts | 🚀 Quick Start | 🎥 Video | ✉️ Contact

---

Domain-RAG is a novel retrieval-augmented generative framework designed for Cross-Domain Few-Shot Object Detection (CD-FSOD). We leverage large-scale vision-language models (GroundingDINO), a curated COCO-style retrieval corpus, and Flux-based background generation to synthesize diverse, domain-aware training data that enhances FSOD generalization under domain shift.

---

✨ Highlights

• 🔍 Retrieval-Augmented Generation: retrieve semantically similar source images for novel-class prompts.
• 🎨 Flux-Redux Integration: compose diverse backgrounds with target foregrounds for domain-aligned generation.
• 📦 Support for Multiple Target Domains: ArTAXOr, Clipart1k, DIOR, DeepFish, UODD, NEU-DET, and more.
• 🧪 Strong Benchmarks: surpasses GroundingDINO baseline in few-shot setting across CD-FSOD & RS-FSOD & CAMO-FS.

---

🔧 Installation

git clone https://github.com/LiYu0524/Domain-RAG.git
cd Domain-RAG
conda create -n domainrag python=3.10
conda activate domainrag
pip install -r requirements.txt

Dataset Preparation

You can prepare CD-FSOD with CDVITO

For NWPU VHR-10(RS-FSOD) dataset, you can download it from NWPU VHR-10

For CAMO-FS dataset, you can download it from CAMO-FS

Quick start

You can refer to ./domainrag.sh

Video

Walkthrough video(Chinese version): Watch here

Contact

For questions and collaboration, please contact:

• Yu Li : <[email protected]>
• Xingyu Qiu : <[email protected]>
• Yuqian Fu : <[email protected]>

Citation

If you find Domain-RAG useful in your research, please cite:

@article{li2025domain,
  title={Domain-RAG: Retrieval-Guided Compositional Image Generation for Cross-Domain Few-Shot Object Detection},
  author={Li, Yu and Qiu, Xingyu and Fu, Yuqian and Chen, Jie and Qian, Tianwen and Zheng, Xu and Paudel, Danda Pani and Fu, Yanwei and Huang, Xuanjing and Van Gool, Luc and others},
  journal={arXiv preprint arXiv:2506.05872},
  year={2025}
}

If you find CD-Vito useful in your research, please cite:

@inproceedings{fu2024cross,
  title={Cross-domain few-shot object detection via enhanced open-set object detector},
  author={Fu, Yuqian and Wang, Yu and Pan, Yixuan and Huai, Lian and Qiu, Xingyu and Shangguan, Zeyu and Liu, Tong and Fu, Yanwei and Van Gool, Luc and Jiang, Xingqun},
  booktitle={European Conference on Computer Vision},
  pages={247--264},
  year={2024},
  organization={Springer}
}