MPG-SAM 2: Adapting SAM 2 with Mask Priors and Global Context for Referring Video Object Segmentation

ICCV 2025

[June 26th, 2025] 🔥MPG-SAM 2 has been accepted by ICCV 2025! We have open-sourced the code and model.

[August 9th, 2025] The Figure 2 in both the ICCV camera-ready version and earlier arXiv versions contains errors. Please refer to the latest version on arXiv for the correct diagram.

• Framework
• Model
• Installation
• Prepare
• Training & Inference
• Acknowledgement
• Citation

Our model can be obtained from this Hugging Face link MPG-SAM 2 · HF. The tiny model uses BEiT-3-base, while the other models use BEiT-3-large. "ytvos" refers to the model trained on the Ref-Youtube-VOS dataset, and "mevis" refers to the model trained on the MeViS dataset.

git clone https://github.com/rongfu-dsb/MPG-SAM2.git
cd MPG-SAM2
conda create -n MPG-SAM2 python=3.8
conda activate MPG-SAM2
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
pip install 'git+https://github.com/facebookresearch/fvcore' 

We mainly used the Ref-Youtube-VOS, Ref-DAVIS17, RefCOCO/+/g, and MeViS datasets. For the data preparation process, please refer to ReferFormer and MeViS. Once the data is prepared, replace ./your_data in the code with the path to your dataset.

├── your_data
│   ├── coco
│   │   ├── refcoco
│   │   ├── refcoco+
│   │   ├── refcocog
│   │   └── train2014
│   ├── ref-youtube-vos
│   │   ├── meta_expressions
│   │   ├── train
│   │   └── valid
│   ├── ref-davis
│   │   ├── DAVIS
│   │   ├── davis_text_annotations
│   │   ├── meta_expressions
│   │   ├── train
│   │   └── valid
│   ├── mevis-release
│   │   ├── train
│   │   ├── valid
│   │   └── valid_u

In this link BEiT-3, download the weights for BEiT-3-large/BEiT-3-base and beit3.spm, and then replace --encoder_pretrained and --version accordingly. In this link SAM2, download the weights for sam2_hiera_large and replace --vision_pretrained.

bash pre_fine.sh

bash ./scripts/dist_test_davis.sh ./workdir_davis ./your_ytvos_weights --ngpu=1

The Ref-DAVIS17 dataset is not used for training; instead, the weights trained on the Ref-Youtube-VOS dataset are directly used for inference.

bash finetune_mevis.sh

The inference results from the Ref-Youtube-VOS and MeViS datasets are submitted to servers Ref-Youtube-VOS and MeViS for metric evaluation, respectively.

We borrow some code from the following works and would like to express our gratitude to them: SAM2, EVF-SAM, ReferFormer, BEiT-3.

@article{rong2025mpg,
  title={MPG-SAM 2: Adapting SAM 2 with Mask Priors and Global Context for Referring Video Object Segmentation},
  author={Rong, Fu and Lan, Meng and Zhang, Qian and Zhang, Lefei},
  journal={arXiv preprint arXiv:2501.13667},
  year={2025}
}