BTSeg: Barlow Twins Regularization for Domain Adaptation in Semantic Segmentation

Paper | Result ACDC

BTSeg employs a novel application of the Barlow Twins loss, a concept borrowed from unsupervised learning. The original Barlow Twins approach uses stochastic augmentations in order to learn useful representations from unlabeled data without the need for external labels. In our approach, we regard images captured at identical locations but under varying adverse conditions as manifold representation of the same scene (which could be interpreted as ”natural augmentations”), thereby enabling the model to conceptualize its understanding of the environment.

---

Official PyTorch implementation of the paper BTSeg: Barlow Twins Regularization for Domain Adaptation in Semantic Segmentation.

The code is organized using PyTorch Lightning.

Setup:

1. Create environment using conda:

 conda env create --file conda_env.yml
 conda activate btseg-env

2. Create an environment file (.env) and specify the location of the datasets and where to save the log files.

 DATA_DIR="/path/to/root/directory/data"
 OUTPUT_DIR="/path/to/output/dir"

3. Download the required pretrained weights.

• SegFormer weights pretrained on Cityscapes (segformer.b5.1024x1024.city.160k.pth)
• UAWarpC pretrained on MegaDepth
• save them in ./pretrained_weights/

4. Download checkpoint

• download the checkpoint of our model used for comparison with the state of the art
• store it in ./checkpoints/

4. Prepare the ACDC data

• Register and download the data from https://acdc.vision.ee.ethz.ch/

• The scripts assume that the folder is named "ACDC" and is located under the path specified in DATA_DIR

• Download the segmentation maps for the reference images and merge them into the ACDC root folder

  • generated using Hierarchical Multi-Scale Attention for Semantic Segmentation

• Final structure:

  • /path/to/ACDC
    • gt
      • fog
        • test_ref
        • train
        • train_ref
        • val
        • val_ref
      • night
      • ...
    • rgb_anon
      • fog
      • night
      • ...

• Run warp_target_images.py to align the target (adverse weather conditions) to the source (reference weather conditions) images

  python warp_target_images.py

• Generates a *_rgb_anon_warped.png for every *_rgb_anan.png image

5. (Optional) Create the ACG benchmark dataset.

• Download the list of filenames and follow the instructions from here
• name the folder "ACG_benchmark"

Test:

On the ACDC validation subset (should result in a mean IoU for the adverse images of 0.69925):

python src/run.py test -c ./configs/Test_BT-Seg_ACDC.yaml --ckpt_path ./checkpoints/BTSeg_ACDC_best.ckpt --data.test_pred_on_val True

Evaluation on rainy images only:

python src/run.py test -c ./configs/Test_BT-Seg_ACDC.yaml --ckpt_path ./checkpoints/BTSeg_ACDC_best.ckpt --data.test_pred_on_val True --data.weather_condition rain

On the ACG benchmark dataset:

python src/run.py test -c ./configs/Test_BT-Seg_ACG.yaml --ckpt_path ./checkpoints/BTSeg_ACDC_best.ckpt

Do prediction on test data:

python src/run.py predict -c ./configs/Test_BT-Seg_ACDC.yaml --ckpt_path ./checkpoints/BTSeg_ACDC_best.ckpt --model.predict_mode png
# model.predict_mode matplotlib is also possible, to generate easy readable segmentation maps + images
python src/run.py predict -c ./configs/Test_BT-Seg_ACDC.yaml --ckpt_path ./checkpoints/BTSeg_ACDC_best.ckpt --data.test_pred_on_val True --model.predict_mode matplotlib

Training:

python src/run.py fit -c ./configs/Train_BT-Seg_ACDC.yaml
# with smaller input images
python src/run.py fit -c ./configs/Train_BT-Seg_ACDC.yaml --data.image_height 512 --data.image_width 512

Citation

If you find this code useful in your research, please consider citing the paper:

@misc{btseg,
      title={BTSeg: Barlow Twins Regularization for Domain Adaptation in Semantic Segmentation},
      author={Johannes K{\"u}nzel and Anna Hilsmann and Peter Eisert},
      year={2023},
      eprint={2308.16819},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

License

This repository is released under the MIT license. However, care should be taken to adopt the appropriate license for third-party code in this repository. Third-party code is marked as such.