Discriminative Perception via Anchored Description for Reasoning Segmentation

CVPR 2026 | [Paper] | [Code]

DPAD is a reinforcement learning framework for reasoning segmentation that actively cultivates discriminative perception:

• 🎯 Contrastive Reward: Anchor-based discrimination.

• 🧠 Focused Reasoning: -42% shorter chains.

• 🌱 High Efficiency: 3K samples & interpretable.

Installation

git clone https://github.com/mrazhou/DPAD.git
cd DPAD

conda create -n dpad python=3.11
conda activate dpad

pip install torch==2.5.1 torchvision==0.20.1
pip install -e .
pip install sam2 matplotlib

Training

bash training_scripts/run_qwen2_5_3b_refCOCOg.sh

Merge Checkpoint (optional)

python3 training_scripts/model_merger.py --local_dir [path_to_your_actor_checkpoint]

Evaluation

bash evaluation_scripts/eval_all.sh [path_to_your_actor_checkpoint]/actor

Note: The current code has been organized to some extent. Feel free to open an issue or contact me via email for updates and maintenance.

Results

Citation

If you find this repository helpful, please consider citing our paper:

@article{zhou2026DPAD, 
  title={Discriminative Perception via Anchored Description for Reasoning Segmentation}, 
  author={Yang, Tao and Zhou, Qing and Wang, Qi}, 
  journal={CVPR}, 
  year={2026},
}
@article{zhou2026rise, 
  title={Reasoning via Implicit Self-supervised Emergence for Instruction Segmentation}, 
  author={Zhou, Qing and Yang, Lichang and Jia, Yuyu and Gao, Junyu and Ni, Weiping and Wu, Junzheng and Wang, Qi}, 
  volume={40}, 
  number={16}, 
  journal={Proceedings of the AAAI Conference on Artificial Intelligence}, 
  year={2026},
  pages={13746-13754}
}

and the Seg-Zero paper:

@article{liu2025segzero,
  title        = {Seg-Zero: Reasoning-Chain Guided  Segmentation via Cognitive Reinforcement},
  author       = {Liu, Yuqi and Peng, Bohao and Zhong, Zhisheng and Yue, Zihao and Lu, Fanbin and Yu, Bei and Jia, Jiaya},
  journal      = {arXiv preprint arXiv:2503.06520},
  year         = {2025}
}

Acknowledgments

Thanks very much to Seg-Zero, Qwen2.5-VL and SAM2 for their great work.