Code for GRSS DFC 2025 track1.

Docker

Alternatively, you could pull the docker image and test.

sudo su
docker pull registry.ap-northeast-1.aliyuncs.com/liuwang20144623/dfc2025track1:v1
docker images
docker run -it --shm-size=60g --gpus all [image_id] /bin/bash
cd /workspace/DFC2025Track1
pip install mmpretrain
bash run_report.sh

---

or testing in local env:

1.Conda env

Cuda version is 11.8.

conda create -n mmseg_dfc python=3.10
source activate mmseg_dfc
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
pip install mmcv=2.1.0 -f https://download.openmmlab.com/mmcv/dist/cu118/torch2.1/index.html
pip install mmengine==0.10.2
cd src/causal-conv1d-1.0.2
pip install -e .
cd ../mamba-1.0.1
pip install -e .
cd ../..
pip install -e .
pip install ttach
pip install kornia
pip install ftfy
pip install scikit-image
pip install timm
pip install mmpretrain

2.Data prepare

Please see in file 'data/DFC2025Track1/copy_dataset_here'.

• If testing only, please copy the OEM-SAR test images to 'data/DFC2025Track1/test/sar_images'.
• If training, the open-earth-map (OEM) and open-erath-map-SAR (OEM-SAR) are utilized to train our method.
  Please reorganize the dir tree as shown in 'data/copy_dataset_here'.

3.Test

Reproduce similarity results: The trained model weights should be downloaded here, password=433w

bash run_report.sh

Note that the final results is slightly different due to the various type of GPUs.

4.Train

You can also find the training pipline for a single model in 'run_pipline.sh':`

bash run_pipline.sh