Pytorch implementation of Non-Local LinkNet (NL-LinkNet) for Multi-class Road Extraction

• Cuda >= 8.0
• Python >= 3.5
• Torchvision >= 0.2.1
• Torch >= 1.0.0
• cv2 >= 3.4.0
• numpy, matplotlib, scikit-image, scipy, pickle, argparse

Install prerequisites with: pip3 install -r requirements.txt

├── road
│   ├── asset
│   │   ├── *.png 
│   │   └── ...
│   ├── label
│   │   ├── *.json
│   │   └── ...
├──────── ...

cd dataset
python3 split_train_val_nia.py --dataset DATASET_PATH --val_frac 0.1`
cd ..

• 학습: CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train_nia.py --model NL34_LinkNet --name nia_cls_7 --crop_size 768 768 --batch_per_gpu 20 --num_cls 7 --total_epoch 500 --dataset dataset/train
• 추론: CUDA_VISIBLE_DEVICES=0 python3 test_nia.py --n_class 7 --name nia_cls_7_500 --model NL34_LinkNet --source dataset/valid
• 평가: python3 eval.py --n_class 7 --name nia_cls_7_500

• 학습: CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train_nia.py --model NL34_LinkNet --name nia_cls_2 --crop_size 768 768 --batch_per_gpu 20 --num_cls 2 --total_epoch 500 --dataset dataset/train
• 추론: CUDA_VISIBLE_DEVICES=0 python3 test_nia.py --n_class 2 --name nia_cls_2_500 --model NL34_LinkNet --source dataset/valid
• 평가: python3 eval.py --n_class 2 --name nia_cls_2_500

• Pre-training Weight: Google Drive
• 7개 클래스 평균 mIOU: 0.385
• background 제외 6개 클래스 평균 mIOU: 0.297

• Pre-trainign Weight: Google Drive
• 2개 클래스 평균 mIOU: 0.782

다음 링크에서 본 소스코드 사용법을 참고할 수 있습니다.

https://www.youtube.com/watch?v=sLSPWgD8w0A

@article{wang2020nl,
  title={NL-LinkNet: Toward Lighter but More Accurate Road Extraction with Non-Local Operations},
  author={Wang, Yooseung and Seo, Junghoon and Jeon, Taegyun},
  journal={IEEE Geoscience and Remote Sensing Letters},
  year={2020}
}