🚀 PF-RPN: Prompt-Free Universal Region Proposal Network (CVPR 2026)

Qihong Tang^1,*, Changhan Liu^1*, Shaofeng Zhang², Wenbin Li¹, Qi Fan^1,📧, Yang Gao¹

* Equal contribution 📧 Corresponding author

¹ Nanjing University, ² University of Science and Technology of China

Official implementation of PF-RPN built on MMDetection.

📌 News

• 2026-03-11: 🧩Codebase and pretrained checkpoint released.
• 2026-02-21: 🎉Our paper has been accepted by CVPR 2026!

✨ Highlights

• ✅ Prompt-free open-set proposal generation with a unified class token (object)
• ✅ Strong AR on both CD-FSOD and ODinW13
• ✅ End-to-end training/evaluation pipeline based on MMDetection
• ✅ One-class annotation generation script for reproducible protocol

🧠 Abstract

Open-vocabulary detectors usually rely on text prompts (class names), which can be unavailable, noisy, or domain-sensitive in deployment. PF-RPN revisits region proposal generation under a prompt-free setting, where all categories are unified into a single token (object).

PF-RPN improves proposal quality with three key designs:

• Sparse Image-Aware Adapter: pseudo text construction from multi-level visual features.
• Cascade Self-Prompt: iterative visual-text enhancement via masked pooling.
• Centerness-Guided Query Selection: top-k decoder query selection using joint confidence.

🏗️ Model Overview

📊 Main Results

PF-RPN achieves state-of-the-art AR under prompt-free evaluation on both CD-FSOD and ODinW13.

Dataset	Method	Prompt Free	AR100	AR300	AR900	ARs	ARm	ARl
CD-FSOD	GDINO†	✗	52.9	53.5	54.7	31.1	41.6	63.9
CD-FSOD	GDINO‡	✓	54.7	57.8	61.6	34.1	49.3	67.0
CD-FSOD	YOLOE-v8-L†	✗	44.4	46.2	47.1	21.6	36.6	54.9
CD-FSOD	YWorldv8-L†	✗	49.6	51.1	51.6	25.1	42.7	60.6
CD-FSOD	Qwen-VL†	✗	20.1	20.1	20.1	1.0	3.0	26.5
CD-FSOD	GLIP†	✗	47.6	47.6	47.6	21.2	34.6	56.0
CD-FSOD	GenerateU	✓	47.7	54.1	55.7	28.1	48.3	69.4
CD-FSOD	Open-Det	✓	36.6	46.3	54.3	28.2	45.3	67.7
CD-FSOD	RPN	✓	32.0	39.0	45.7	29.9	43.0	54.3
CD-FSOD	Cascade RPN	✓	45.8	52.0	56.9	31.1	50.5	66.0
CD-FSOD	PF-RPN (Ours)	✓	60.7	65.3	68.2	38.5	61.9	80.3
ODinW13	GDINO†	✗	72.1	73.4	74.0	45.6	61.7	79.2
ODinW13	GDINO‡	✓	69.1	70.9	72.4	40.8	64.6	78.4
ODinW13	YOLOE-v8-L†	✗	66.6	67.8	68.3	39.2	57.8	72.8
ODinW13	YWorldv8-L†	✗	69.1	70.3	71.5	37.5	62.2	75.4
ODinW13	GLIP†	✗	69.8	69.8	69.8	33.2	50.9	75.2
ODinW13	GenerateU	✓	67.3	71.5	72.2	32.8	63.1	80.0
ODinW13	Open-Det	✓	53.9	62.9	69.1	27.7	59.8	76.6
ODinW13	RPN	✓	49.0	52.4	55.7	35.3	54.0	59.8
ODinW13	Cascade RPN	✓	60.9	65.5	70.2	40.3	65.5	75.0
ODinW13	PF-RPN (Ours)	✓	76.5	78.6	79.8	45.4	71.9	85.8

• † uses original class names as text prompts.
• ‡ replaces class names with object (prompt-free setting).

🧩 Release Status

• Training / evaluation code
• PF-RPN checkpoint
• Data preprocessing utility (tools/merge_classes_and_sample_subset.py)

⚙️ Installation

Validated environment:

• Python 3.10
• CUDA 11.8
• PyTorch 2.1.0

# 1) Create environment
conda create -n pf-rpn python=3.10 -y
conda activate pf-rpn

# 2) Install PyTorch
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 \
  --index-url https://download.pytorch.org/whl/cu118

# 3) Install MMEngine / MMCV
pip install mmengine
pip install "mmcv>=2.0.0" \
  -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.1.0/index.html

# 4) Install this repo
pip install "setuptools>=69.0.3,<81"
pip install -v -e . --no-build-isolation

# 5) Install extras
pip install -r requirements.txt

# 6) Keep NumPy in 1.x line for compatibility
pip install "numpy<2"

⚡ Quick Start

1) Download checkpoints

mkdir -p checkpoints

wget -O checkpoints/groundingdino_swinb_cogcoor_mmdet-55949c9c.pth \
  https://download.openmmlab.com/mmdetection/v3.0/grounding_dino/groundingdino_swinb_cogcoor_mmdet-55949c9c.pth

wget -O checkpoints/pf_rpn_swinb_5p_coco_imagenet.pth \
  https://huggingface.co/tangqh/PF-RPN/resolve/main/pf_rpn_swinb_5p_coco_imagenet.pth

2) One-command evaluation

python tools/test.py \
  configs/pf-rpn/pf-rpn_coco-imagenet.py \
  checkpoints/pf_rpn_swinb_5p_coco_imagenet.pth

🗂️ Data Preparation

Datasets are not bundled in this repository. Please prepare all data under data/.

A) Source-domain training data (COCO 2017 + ImageNet)

The paper describes the source-domain training data as 5% of COCO and 5% of ImageNet. For exact reproduction, however, do not use a direct download of the original ImageNet-1k dataset. Our training used:

• a 5% subset of COCO 2017; and
• a 5% subset of the training set released by the Few-Shot Object Detection Dataset (FSOD), whose images and annotations are available from the official Google Drive folder.

The FSOD release contains the ImageNet-derived training data used in our experiments because original ImageNet-1k is only a classification dataset instead of an object detection one.

For the most reliable reproduction, download the released merged training annotation directly:

• merged_one_class_area.json

It already contains the selected 5% COCO and 5% FSOD-release training samples, with all categories merged into the single object class.

The released config (configs/pf-rpn/pf-rpn_coco-imagenet.py) expects:

data/coco/
  train2017/
    <COCO training images>
    part_1/
    part_2/
  val2017/
    <COCO validation images>
  annotations/
    merged_one_class_area.json
    instances_val2017_sc.json
    instances_val2017_1p_sc.json

Keep the FSOD image subdirectories (including part_1/ and part_2/) under data/coco/train2017/. Do not flatten or rename them: the file_name entries in merged_one_class_area.json retain these relative paths, and the training config resolves them relative to data/coco/train2017/.

Generate one-class annotations from COCO JSON (optional):

Note: The 5% COCO command below only creates a standalone COCO subset. It does not create the final combined COCO+ImageNet training annotation used by the released checkpoint. Use the released merged_one_class_area.json for exact reproduction.

# 5% training subset + merge all categories into one class
python tools/merge_classes_and_sample_subset.py \
  --input data/coco/annotations/instances_train2017.json \
  --output data/coco/annotations/instances_train2017_5p_sc.json \
  --subset-ratio 0.05 \
  --seed 42 \
  --merge-categories

# Full validation split + merge categories
python tools/merge_classes_and_sample_subset.py \
  --input data/coco/annotations/instances_val2017.json \
  --output data/coco/annotations/instances_val2017_1p_sc.json \
  --subset-ratio 1.0 \
  --seed 42 \
  --merge-categories

# Keep config-compatible val filename
cp data/coco/annotations/instances_val2017_1p_sc.json \
  data/coco/annotations/instances_val2017_sc.json

B) CD-FSOD (6 targets)

Benchmark: https://github.com/lovelyqian/CDFSOD-benchmark

Expected structure:

data/cdfsod/
  ArTaxOr/
    test/
    annotations/test_one_class.json
  clipart1k/
    test/
    annotations/test_one_class.json
  DIOR/
    test/
    annotations/test_one_class.json
  FISH/
    test/
    annotations/test_one_class.json
  NEUDET/
    test/
    annotations/test_one_class.json
  UODD/
    test/
    annotations/test_one_class.json

Generate one-class test annotations:

for d in ArTaxOr clipart1k DIOR FISH NEUDET UODD; do
  python tools/merge_classes_and_sample_subset.py \
    --input data/cdfsod/${d}/annotations/test.json \
    --output data/cdfsod/${d}/annotations/test_one_class.json \
    --subset-ratio 1.0 \
    --seed 42 \
    --merge-categories
done

C) ODinW13

Benchmark reference: https://github.com/microsoft/GLIP#the-object-detection-in-the-wild-benchmark

Expected root (matches configs/pf-rpn/ODinW13/*.py):

data/odinw/
  AerialMaritimeDrone/large/
  Aquarium/Aquarium Combined.v2-raw-1024.coco/
  CottontailRabbits/
  EgoHands/generic/
  NorthAmericaMushrooms/North American Mushrooms.v1-416x416.coco/
  Packages/Raw/
  PascalVOC/
  Raccoon/Raccoon.v2-raw.coco/
  ShellfishOpenImages/raw/
  VehiclesOpenImages/416x416/
  pistols/export/
  pothole/
  thermalDogsAndPeople/

Generate one-class validation annotations (non-pistols):

python tools/merge_classes_and_sample_subset.py \
  --input data/odinw/<subset>/valid/annotations.json \
  --output data/odinw/<subset>/valid/annotations_one_class.json \
  --subset-ratio 1.0 \
  --seed 42 \
  --merge-categories

pistols special case:

python tools/merge_classes_and_sample_subset.py \
  --input data/odinw/pistols/export/val_annotations.json \
  --output data/odinw/pistols/export/annotations_one_class.json \
  --subset-ratio 1.0 \
  --seed 42 \
  --merge-categories

🏋️ Training

Single GPU

python tools/train.py configs/pf-rpn/pf-rpn_coco-imagenet.py

Multi-GPU (DDP)

bash tools/dist_train.sh configs/pf-rpn/pf-rpn_coco-imagenet.py 8

Full schedule override

The released config sets train_cfg.max_epochs=1 for quick sanity checks. For a paper-style schedule, override at launch time:

python tools/train.py configs/pf-rpn/pf-rpn_coco-imagenet.py \
  --cfg-options train_cfg.max_epochs=12

🧪 Evaluation

Main config

python tools/test.py \
  configs/pf-rpn/pf-rpn_coco-imagenet.py \
  checkpoints/pf_rpn_swinb_5p_coco_imagenet.pth

CD-FSOD (all 6 configs)

for cfg in configs/pf-rpn/CDFSOD/*.py; do
  python tools/test.py "$cfg" checkpoints/pf_rpn_swinb_5p_coco_imagenet.pth
done

ODinW13 (all 13 configs)

for cfg in configs/pf-rpn/ODinW13/*.py; do
  python tools/test.py "$cfg" checkpoints/pf_rpn_swinb_5p_coco_imagenet.pth
done

Distributed evaluation

bash tools/dist_test.sh \
  configs/pf-rpn/pf-rpn_coco-imagenet.py \
  checkpoints/pf_rpn_swinb_5p_coco_imagenet.pth \
  8

📐 Prompt-Free Protocol

This repository follows a strict one-class open-set setup:

• custom_classes = ('object',)
• Category IDs are merged into one class via tools/merge_classes_and_sample_subset.py
• Evaluation configs in configs/pf-rpn/CDFSOD and configs/pf-rpn/ODinW13 assume one-class annotations

✅ Reproducibility Checklist

Before reporting numbers, verify:

• Environment matches (Python 3.10, PyTorch 2.1.0, CUDA 11.8)
• Checkpoint files are placed in checkpoints/
• COCO 2017 and the FSOD release ImageNet-derived training data are both downloaded
• FSOD image paths such as part_1/ and part_2/ are preserved under data/coco/train2017/
• merged_one_class_area.json is placed at the path expected by the training config
• One-class annotation JSON files are generated and paths match config files
• custom_classes=('object',) is preserved
• Evaluation uses the provided benchmark-specific configs

📚 Citation

If you find PF-RPN is useful in your research or applications, please consider giving us a star 🌟 and citing it.

@article{tang2026prompt,
  title={Prompt-Free Universal Region Proposal Network},
  author={Tang, Qihong and Liu, Changhan and Zhang, Shaofeng and Li, Wenbin and Fan, Qi and Gao, Yang},
  journal={arXiv preprint arXiv:2603.17554},
  year={2026}
}

🙏 Acknowledgement

This project is built upon:

• MMDetection
• GroundingDINO
• GLIP and ODinW benchmark resources

We also thank the FSOD authors for releasing the ImageNet-derived training data used in our experiments.

• Few-Shot-Object-Detection-Dataset