Research map / Detection and segmentation / Open-set and few-shot detection
Domain-RAG: Retrieval-Guided Compositional Image Generation for Cross-Domain Few-Shot Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Domain-RAG, a training-free retrieval-guided compositional image generation pipeline (LaMa inpainting, CLIP+ResNet-style retrieval from COCO, FLUX-Redux/FLUX-Fill) that augments support sets for cross-domain few-shot object detection with fine-tuned GroundingDINO.
Problem
Cross-domain few-shot object detection requires few labeled target samples, and existing augmentation or text-to-image generation methods often fail to preserve object categories or produce backgrounds consistent with the target domain.
Approach
Keep the annotated foreground fixed and adapt the background: retrieve semantically (CLIP) and stylistically (shallow ResNet statistics) similar COCO backgrounds, fuse their embeddings with the inpainted target background via FLUX-Redux to generate a domain-aligned background, then recompose the foreground with FLUX-Fill, all without training.
Key results
- On CD-FSOD, Domain-RAG reaches 33.6 average mAP in 1-shot vs 26.3 for fine-tuned GroundingDINO and 27.8 for ETS, with 42.7 (5-shot) and 45.4 (10-shot) averages.
- On NWPU VHR-10 RS-FSOD, it improves SEA-FSDet from 68.36 to 70.67 average mAP across 3/5/10/20-shot.
- On CAMO-FS, Domain-RAG averages 68.0 mAP vs 66.6 for the GroundingDINO baseline across 1/2/3/5-shot.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Domain-RAG augmentation to GroundingDINO | GroundingDINO on CD-FSOD · CD-FSOD 1-shot, six-domain average | mAP: 26.3 -> 33.6 | Helped |
| add Domain-RAG augmentation to GroundingDINO | GroundingDINO on CD-FSOD · CD-FSOD 5-shot, six-domain average | mAP: 40.4 -> 42.7 | Helped |
| add Domain-RAG augmentation to GroundingDINO | GroundingDINO on CD-FSOD · CD-FSOD 10-shot, six-domain average | mAP: 43.9 -> 45.4 | Helped |
| add Domain-RAG augmentation to SEA-FSDet | SEA-FSDet on NWPU VHR-10 RS-FSOD · NWPU VHR-10, average over 3/5/10/20-shot | mAP: +2.31 mAP | Helped |
| add Domain-RAG augmentation to SEA-FSDet | SEA-FSDet on NWPU VHR-10 RS-FSOD · NWPU VHR-10, 20-shot | mAP: 85.08 to 84.05 | Hurt |
| add Domain-RAG augmentation to GroundingDINO | GroundingDINO on CAMO-FS · CAMO-FS, average over 1/2/3/5-shot | mAP: 66.6 -> 68.0 | Helped |
| add domain-aware background retrieval (ablation) | Domain-RAG on CD-FSOD · CD-FSOD 1-shot, six-domain average | mAP: 31.8 -> 33.6 | Helped |
| add domain-guided background generation (ablation) | Domain-RAG on CD-FSOD · CD-FSOD 1-shot, six-domain average | mAP: 31.1 -> 33.6 | Helped |
| use generative foreground-background composition instead of copy-paste (ablation) | Domain-RAG on CD-FSOD · CD-FSOD 1-shot, six-domain average | mAP: 29.2 -> 33.6 | Helped |
| use COCO-80classes rather than a reduced-category COCO database (ablation) | Domain-RAG on CD-FSOD · CD-FSOD 1-shot, six-domain average | mAP: 30.9 -> 33.6; 31.5 -> 33.6; 32.3 -> 33.6 | Helped |
| combine CLIP semantic retrieval with ResNet style retrieval (ablation) | Domain-RAG on CD-FSOD · CD-FSOD 1-shot, six-domain average | mAP: 31.5 -> 33.6; 30.7 -> 33.6 | Helped |
| retrieve m = 5 images rather than m = 1 or m = 3 (ablation) | Domain-RAG on CD-FSOD · CD-FSOD 1-shot, six-domain average | mAP: 29.2 -> 33.6; 31.4 -> 33.6 | Helped |
Surprising findings
- Removing the background retrieval module or the compositional stage still leaves gains, but the 20-shot NWPU result slightly decreases (85.08 to 84.05), which the authors attribute to overfitting on synthetic patterns when base training is sufficient.
- CLIP-I and FID are argued to be unreliable for judging generation quality in cross-domain settings, so higher CLIP-I or lower FID does not necessarily mean better augmentation.
Limitations
- Foreground information leakage: parts of the foreground are regenerated in the background without annotations, which may add noise to fine-tuning.
- Relies on LaMa inpainting quality; large foregrounds produce artifacts that Redux propagates.
- No effective filtering mechanism; general-purpose CLIP fails to filter poor generated samples in cross-domain settings.
Experimental setup
Datasets: COCO ArTaxOr Clipart1k DIOR DeepFish NEU-DET UODD NWPU VHR-10 CAMO-FS miniImageNet
Benchmarks: CD-FSOD (CD-ViTO benchmark: ArTaxOr, Clipart1k, DIOR, DeepFish, NEU-DET, UODD) NWPU VHR-10 (RS-FSOD and CD-FSOD setups) CAMO-FS
Metrics: mAP CLIP-I FID
Baselines: GroundingDINO ETS CD-ViTO Distill-cdfsod DE-ViT Detic-FT Copy-Paste Background Augmentation Foreground Augmentation SEA-FSDet
Builds on: GroundingDINO ETS FLUX FLUX-Redux FLUX-Fill LaMa
Closest papers by meaning
- Fine-Grained Prototypes Distillation for Few-Shot Object Detection (AAAI 2024 · similarity 0.88)
- Multi-Perspective Data Augmentation for Few-shot Object Detection (ICLR 2025 · similarity 0.87)
- Revisiting Few-Shot Object Detection with Vision-Language Models (NeurIPS 2024 · similarity 0.86)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.86)
- Non-exemplar Domain Incremental Object Detection via Learning Domain Bias (AAAI 2024 · similarity 0.86)
- Weakly Supervised Few-Shot Object Detection with DETR (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.