Beyond Prompt Degradation: Prototype-guided Dual-pool Prompting for Incremental Object Detection

🎉 Accepted at CVPR 2026 🎉

Official implementation of PDP: Prototype-guided Dual-pool Prompting for Incremental Object Detection

---

🔥 Introduction

Incremental Object Detection (IOD) aims to continuously learn new object categories without forgetting previously learned ones.

Recent prompt-based IOD methods are replay-free and parameter-efficient. However, they suffer from two critical issues:

• Prompt Coupling: Task-general and task-specific prompts interfere with each other.
• Prompt Drift: Inconsistent supervision causes old categories to degrade.

To address these challenges, we propose PDP, a prompt-decoupled continual detection framework that explicitly separates transferable and task-specific knowledge while maintaining supervision consistency across incremental steps.

---

🧠 Key Contributions

1️⃣ Dual-Pool Prompt Decoupling

• Shared Prompt Pool: captures task-general transferable knowledge
• Private Prompt Pool: learns task-specific discriminative features
• Explicit decoupling mitigates prompt interference

2️⃣ Prototypical Pseudo-Label Generation (PPG)

• Dynamically maintains class prototype space
• Filters teacher-generated pseudo labels using prototype similarity
• Ensures supervision consistency during incremental training

---

📊 Results

Dataset	Setting	Performance Gain
MS-COCO	21+19+20+20	+9.2% AP
PASCAL VOC	19+1	+3.3% AP

PDP achieves state-of-the-art performance while remaining replay-free and parameter-efficient.

---

🏗️ Installation

git clone https://github.com/zyt95579/PDP_IOD
cd PDP_IOD
conda create -n pdp python=3.8 -y
conda activate pdp
pip install -r requirement.txt

🚀 Usage

🏋️ Training

bash run.sh

🏋️ Testing

To evaluate the trained model on the test set, first modify the run.sh script: Change the parameter train=1 to train=0 Then run the same command:

bash run.sh

📝 Citation

If you find this repo useful, please cite:

@inproceedings{zhang2026beyond,
  title={Beyond Prompt Degradation: Prototype-guided Dual-pool Prompting for Incremental Object Detection},
  author={Zhang, Yaoteng and Zhou, Qing and Gao, Junyu and Wang, Qi},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={27568--27578},
  year={2026}
}