🌳 Code Structure

├─MVLPNet
|   ├─utils.py
|   ├─vis.py # vis.py is the code for visualization
|   ├─test.py
|   ├─test.sh
|   ├─train.py
|   ├─train.sh
|   ├─train_base.py
|   ├─train_base.sh
|   ├─util
|   ├─model
|   |   ├─workdir
|   |   ├─util
|   |   ├─few_seg
|   |   |    └MVLPNet.py
|   |   ├─backbone
|   |   ├─clip
|   ├─lists
|   |   ├─iSAID
|   |   ├─LoveDA
|   |   ├─iSAID_256
|   ├─initmodel
|   |     ├─PSPNet
|   |     ├─CLIP
|   ├─vgg16_bn.pth
|   ├─resnet50_v2.pth
|   ├─exp
|   ├─dataset
|   ├─config
├─data
|  ├─iSAID
|  |   ├─train.txt
|  |   ├─val.txt
|  |   ├─img_dir
|  |   ├─ann_dir

📝 Data Preparation

• Create a folder data at the same level as this repo in the root directory.

  cd ..
  mkdir data
  

• iSAID_512: iSAID.tar.gz : https://pan.baidu.com/s/11ZhZ01KVjfPyHcoZ2MkfeA password: 2025

• iSAID_256: iSAID.tar.gz : https://pan.baidu.com/s/1WgZBH075gjmypS4NbiLaXg password: 2025

• LoveDA: LoveDA.tar.gz : https://pan.baidu.com/s/1XG7zsh5uTOerffrE73cj2g password: 2025

Train

Training base-learners (two options)

• Option 1: training from scratch

  Download the pre-trained backbones from (https://pan.baidu.com/s/1tWAUKYvP-sh_LcCOy1-P7Q password: 2025) and put them into the MVLPNet/initmodel directory. The clip model is placed in the MVLPNet/initmodel directory: (https://pan.baidu.com/s/1vwtIinePOP7UdhrEDj4HKg password: 2025)

  sh train_base.sh
  

• Option 2: loading the trained models

  mkdir initmodel
  cd initmodel
  

  Put the provided (https://pan.baidu.com/s/1I4s8PLy4N5Qb7UeE7VsVXQ password: 2025) in the newly created folder initmodel and rename the downloaded file to PSPNet, i.e., MVLPNet/initmodel/PSPNet.

Training few-shot models

To train a model, run

sh train.sh

Testing few-shot models

To evaluate the trained models, run

sh test.sh

👏 Acknowledgements

The project is based on PFENet , R2Net, DMNet and PI-CLIP. Thanks for the authors for their efforts.

📄 Citation

Please cite our paper if you find it is useful for your research.

@ARTICLE{11071646,
  author={Yang, Zhenhao and Bi, Fukun and Han, Jianhong and Ma, Xianping and He, Chenglong and Liu, Wenkai},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Multimodal Visual-Language Prompt Network for Remote Sensing Few-Shot Segmentation}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  keywords={Few-shot segmentation;remote sensing;text prompts;semantic segmentation},
  doi={10.1109/TGRS.2025.3585878}}