Self-Pair: Synthesizing Changes from Single Source for Object Change Detection in Remote Sensing Imagery (WACV2023)

Update Note

• 2022.12.22 : Update VSAIT style transfer method
• This repository will be completed on the date of the WACV 2023 presentation.

Install & Train (VSAIT-style transfer)

Sample

prepare

cd vsait
pip install -r requirements.txt

|./data/
|------/source
|-------------/train/A
|-------------/val/A
|------/target
|-------------/train/B
|-------------/val/B

python train.py --name="vsait"
python test.py --name="vsait_adapt" --checkpoint="./checkpoints/vsait/version_0/checkpoints/epoch={i}-step={j}.ckpt"

Install (Open-CD)

pip install -U openmim
mim install mmcv-full

git clone https://github.com/seominseok0429/Self-Pair-for-Change-Detection.git
cd open-cd
pip install -v -e .

train

python tools/train.py configs/changer/changer_ex_r18_512x512_40k_levircd.py --work-dir ./changer_r18_levir_workdir --gpu-id 0 --seed 307

Results

Method	Backbone	Dataset	Param (M)	MACs (G)	Precision	Recall	F1	config
Changer (ChangerEx)	ResNet18	LEVIR-CD	11.39	23.71	92.24	91.09	91.66	changer_ex_r18_512x512_40k_levircd.py
Changer + Self-pair (ChangerEx)	ResNet18	LEVIR-CD	11.39	23.71	TODO	TODO	TODO	TODO

Citation

If you find this project useful in your research, please consider cite:

@article{seo2023selfpair,
  title={Self-Pair: Synthesizing Changes from Single Source for Object Change Detection in Remote Sensing Imagery}, 
  author={Minseok Seo, Hakjin Lee, Yongjin Jeon, Junghoon Seo},
  year={2023},
  eprint={2212.10236},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

@article{fang2022changer,
  title={Changer: Feature Interaction is What You Need for Change Detection}, 
  author={Sheng Fang and Kaiyu Li and Zhe Li},
  year={2022},
  eprint={2209.08290},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}


@misc{opencd2022,
    title={{Open-CD}: An open source change detection toolbox},
    author={Open-CD Contributors},
    howpublished = {\url{https://github.com/likyoo/open-cd}},
    year={2022}
}

Thank to

This code is heavily based on open-cd.

We thank the authors of that code.