---
license: apache-2.0
tags:
- object-detection
- region-proposal
- open-set-detection
- zero-shot-detection
- mmdetection
- pytorch
- cvpr2026
datasets:
- coco
- imagenet
- cd-fsod
- odinw
metrics:
- average-recall (AR)
---

# PF-RPN: Prompt-Free Universal Region Proposal Network

## 🧠 Model Details

**PF-RPN** (Prompt-Free Universal Region Proposal Network) is a state-of-the-art model for Cross-Domain Open-Set Region Proposal Network, accepted at **CVPR 2026**. 

Open-vocabulary detectors typically rely on text prompts (class names), which can be unavailable, noisy, or domain-sensitive during deployment. PF-RPN tackles this by revisiting region proposal generation under a strictly **prompt-free** setting. Instead of specific category names, all categories are unified into a single token (`object`). 

### Model Architecture Innovations
To improve proposal quality without explicit class prompts, PF-RPN introduces three key designs:
1. **Sparse Image-Aware Adapter:** Constructs pseudo-text representations from multi-level visual features.
2. **Cascade Self-Prompt:** Iteratively enhances visual-text alignments via masked pooling.
3. **Centerness-Guided Query Selection:** Selects top-k decoder queries using joint confidence scores.

### Model Sources
- **Repository:** [PF-RPN GitHub Repository](https://github.com/tangqh03/PF-RPN)
- **Paper:** PF-RPN: Prompt-Free Universal Region Proposal Network (CVPR 2026)
- **Base Framework:** [MMDetection 3.3.0](https://github.com/open-mmlab/mmdetection)
- **Backbone:** Swin-Base (`swinb`)

## 🎯 Intended Use

- **Primary Use Case:** Generating high-quality, class-agnostic region proposals ("objects") across diverse, unseen domains without requiring domain-specific text prompts or retraining.
- **Protocol:** Strict one-class open-set setup where `custom_classes = ('object',)`.

## 🗂️ Training Data

The provided checkpoint (`pf_rpn_swinb_5p_coco_imagenet.pth`) was trained on a combined dataset of **COCO 2017** and **ImageNet-1k**. 
- To simulate the open-set proposal generation task, all ground-truth categories are merged into a single class (`object`).
- The specific released model uses a **5% subset** of the COCO training data merged with ImageNet images.

## 📊 Evaluation Data and Performance

PF-RPN achieves state-of-the-art Average Recall (AR) under prompt-free evaluation across multiple benchmarks. 

### Cross-Domain Few-Shot Object Detection (CD-FSOD)
Evaluated across 6 target domains (ArTaxOr, clipart1k, DIOR, FISH, NEUDET, UODD).

| Method | Prompt Free | AR100 | AR300 | AR900 | ARs | ARm | ARl |
|---|:---:|---:|---:|---:|---:|---:|---:|
| GDINO‡ | ✓ | 54.7 | 57.8 | 61.6 | 34.1 | 49.3 | 67.0 |
| GenerateU | ✓ | 47.7 | 54.1 | 55.7 | 28.1 | 48.3 | 69.4 |
| Cascade RPN | ✓ | 45.8 | 52.0 | 56.9 | 31.1 | 50.5 | 66.0 |
| **PF-RPN (Ours)** | **✓** | **60.7** | **65.3** | **68.2** | **38.5** | **61.9** | **80.3** |

### Object Detection in the Wild (ODinW13)
Evaluated across 13 diverse target domains.

| Method | Prompt Free | AR100 | AR300 | AR900 | ARs | ARm | ARl |
|---|:---:|---:|---:|---:|---:|---:|---:|
| GDINO‡ | ✓ | 69.1 | 70.9 | 72.4 | 40.8 | 64.6 | 78.4 |
| GenerateU | ✓ | 67.3 | 71.5 | 72.2 | 32.8 | 63.1 | 80.0 |
| Cascade RPN | ✓ | 60.9 | 65.5 | 70.2 | 40.3 | 65.5 | 75.0 |
| **PF-RPN (Ours)** | **✓** | **76.5** | **78.6** | **79.8** | 45.4 | **71.9** | **85.8** |

*(‡ indicates models where original class names were replaced with `object` to simulate a prompt-free setting).*

## ⚙️ How to Use

### Installation
Ensure you have a working environment with Python 3.10, PyTorch 2.1.0, and CUDA 11.8. Install MMDetection and this repository's codebase as described in the [GitHub README](https://github.com/tangqh03/PF-RPN#%EF%B8%8F-installation).

### Quick Start: Evaluation

1. **Download the Weights**
```bash
mkdir -p checkpoints

# Download GroundingDINO base weights
wget -O checkpoints/groundingdino_swinb_cogcoor_mmdet-55949c9c.pth \
  [https://download.openmmlab.com/mmdetection/v3.0/grounding_dino/groundingdino_swinb_cogcoor_mmdet-55949c9c.pth](https://download.openmmlab.com/mmdetection/v3.0/grounding_dino/groundingdino_swinb_cogcoor_mmdet-55949c9c.pth)

# Download PF-RPN weights
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](https://huggingface.co/tangqh/PF-RPN/resolve/main/pf_rpn_swinb_5p_coco_imagenet.pth)

2. **Run Inference / Testing**
```bash
python tools/test.py \
  configs/pf-rpn/pf-rpn_coco-imagenet.py \
  checkpoints/pf_rpn_swinb_5p_coco_imagenet.pth
```
Note: Data preprocessing is required before evaluation. Datasets must be downloaded and their categories merged into a single `object` class using the provided `tools/merge_classes_and_sample_subset.py` script. See the repository for detailed data preparation commands.

## 📚 Citation
If you use PF-RPN in your research, please cite:
```bibtex
@inproceedings{tang2026pf,
  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},
  booktitle={CVPR},
  year={2026}
}
```