cyx669521 / Efficient-Transformer

Online !!! Application of an efficient transformer improved based on Swin transformer on remote sensing segmentation

Efficient-Transformer-for-Remote-Sensing-Image-Segmentation, RemoteSensing

Introduction

We propose an Efficient Transformer combined with an efficient backbone (ref to Rest) and Mlphead to achieve light-weight design. We also improve the edge segmentation by applying explicit and implicit edge enhancement methods. Finally, our model achieves new state-of-the-art performances on Potsdam and Vaihingen datasets.

Usage

• Install packages

  This repository is based on python 3.6.12 and torch 1.6.0.

  git clone https://github.com/zyxu1996/Efficient-Transformer.git
  cd Efficient-Transformer  
  

  pip install -r requirements.txt
  

• Prepare datasets and pretrained weights

  • Download Potsdam and Vaihingen datasets form BaiduYun, and put them on ./data
BaiduYun: https://pan.baidu.com/s/1MyDw1qncPKYJFK_zjFxFBA
    Password: s7f2

    Data file structure of Potsdam and Vaihingen datasets is as followed.

    ├── data                      ├── data
        ├──potsdam                    ├──vaihingen
           ├──images                     ├──images
              ├──top_potsdam_2_10.tif       ├──top_mosaic_09cm_area1.tif
              ├──top_potsdam_2_11.tif       ├──top_mosaic_09cm_area2.tif
              ...                           ...
           ├──annotations                ├──annotations
              ├──train.txt                  ├──train.txt
              ├──test.txt                   ├──test.txt
              ├──labels                     ├──labels
                 ├──top_potsdam_2_10.png       ├──top_mosaic_09cm_area1.png
                 ├──top_potsdam_2_11.png       ├──top_mosaic_09cm_area2.png
                 ...                           ...
    

  • Download the pretained weights from Swin and Rest, and put them on ./pretrained_weights
    Swin: ImageNet-1K, ImageNet-22K and ADE20K Semantic Segmentation pretrained models are used.
    Rest: ImageNet-1K pretrained models are used.

• Training

  • The training and testing settings are written in the script, including the selection of datasets and models.

    sh autorun.sh
    

  • If directly run train.py, please undo the following code.

    if __name__ == '__main__':
      os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
      os.environ.setdefault('RANK', '0')
      os.environ.setdefault('WORLD_SIZE', '1')
      os.environ.setdefault('MASTER_ADDR', '127.0.0.1')
      os.environ.setdefault('MASTER_PORT', '29556')
    

• Testing

  • Generating the final results and visulizing the prediction.

    cd ./work_dir/your_work
    

  • Do remember undo the test command in sh autorun.sh. And change the --information num1 of testing command as the same as the information in training command.
    CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 --master_port 29506 test.py --dataset vaihingen --val_batchsize 16 --models swinT --head mlphead --crop_size 512 512 --save_dir work_dir --base_dir ../../ --information num1
  • Then run the script autorun.sh.

    sh autorun.sh
    

Citation

@article{xu2021efficient,
  title={Efficient Transformer for Remote Sensing Image Segmentation},
  author={Xu, Zhiyong and Zhang, Weicun and Zhang, Tianxiang and Yang, Zhifang and Li, Jiangyun},
  journal={Remote Sensing},
  volume={13},
  number={18},
  pages={3585},
  year={2021},
  publisher={Multidisciplinary Digital Publishing Institute}
}