CutMix-CD: Advancing Semi-Supervised Change Detection via Mixed Sample Consistency

This repocitory contains the official implementation of our paper: CutMix-CD: Advancing Semi-Supervised Change Detection via Mixed Sample Consistency.

<img width="1179" alt="image" src="https://github.com/user-attachments/assets/717388c3-c9b3-4659-a5b7-ba29cc6c47de" />

:speech_balloon: Requirements

This repo was tested with python 3.8, torch 1.7.1.

pip install -r requirements.txt

:speech_balloon: Data preparation

Download LEVIR-CD and S2Looking datasets. The data preprocessing is introduced in the paper.

Check the file loaders/datasets.py and you can adjust it appropriately for your own dataset.

Modify the argument --data_root in training scripts.

:speech_balloon: Training

We provide training scripts on LEVIR-CD and S2looking datasets.

To train the model, first download ImageNet-pretrained 3x3resnet50-imagenet.pth file and save it to the path models/backbones/pretrained.

Detailed training arguments are described in the training script. You can simply train a model by:

python train_LEVIR.py

During the training, the losses and metrics are reported in the file train.log saved in --log. The best model is saved as best.pth.

:speech_balloon: Inference

Set the path --weight_path of the best model in inference script and evaluate on the test set by:

python inference_LEVIR.py

Evaluation metrics are saved in the file test.log saved in --log.

:speech_balloon: Citation

If you find this repo useful for your research, please consider citing the paper as follows:

@ARTICLE{10810476,
  author={Shu, Qidi and Zhu, Xiaolin and Wan, Luoma and Zhao, Shuheng and Liu, Denghong and Peng, Longkang and Chen, Xiaobei},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={CutMix-CD: Advancing Semi-Supervised Change Detection via Mixed Sample Consistency}, 
  year={2024},
  doi={10.1109/TGRS.2024.3520630}}

Acknowledgements

Thanks to the following open source efforts:

• cutmix-semisup-seg.
• SemiCD.
• FPA-SSCD.