CrackUDA: Incremental Unsupervised Domain Adaptation for Improved Crack Segmentation in Civil Structures.

[Project Page] [Paper]

In Proceedings of the International Conference on Pattern Recognition (ICPR), 2024.

Installation

conda create -n crackuda python=3.9
conda activate crackuda
conda install -c pytorch pytorch=1.10.1 torchvision=0.11.2 cudatoolkit=11.3.1
pip install matplotlib numpy==1.24 tensorboardX opencv-python pillow

Datasets

• CrackSeg9k Used as the source dataset for conducting the experiments reported in the paper.
• BuildCrack: Dataset introduced in this paper, used as target dataset along with sub-datasets of CrackSeg9k.

The expected data sturcture for any dataset is as follows:

   workspace/
   ├── source_dataset (CrackSeg9k) /
   |   ├──train 
   │        ├── images
   |            | image1.jpg
   |            | image2.jpg
   │        ├── labels
   |            | image1.png
   |            | image2.png
   |   ├──val 
   │        ├── images
   |            | image1.jpg
   |            | image2.jpg
   │        ├── labels
   |            | image1.png
   |            | image2.png
   ├── target_dataset (BuildCrack) /
   │   ├── images
   |        | image1.jpg
   |        | image2.jpg
   │   ├── labels
   |        | image1.png
   |        | image2.png

Training

Our approach operates in an incremental learning setting. In Step 1, we train for binary segmentaion on the CrackSeg9k dataset. You may chose to avoid sub-datasets of CrackSeg9k, and use it as a target dataset in Step 2 to perform Unsupervised Domain Adaptation.

python train.py --save_dir <step1_save_path> --num_epochs 150 --source_dataset_path <source_dataset> --dataset_avoided <sub_dataset_name>

CrackSeg9k contains several subdatasets such as Mason, Ceramic, CFD, etc.. --dataset_avoided May or may not be set using the names of these sub-datasets. Use 'NA' if no sub-dataset is to be avoided.

For Step 2, we use the trained model from Step 1 to perform UDA using adversarial training through GRL.

python train_step2.py --save_dir <step2_save_path> --num_epochs 20 --target_dataset_path <target_dataset> --source_dataset_path <source_dataset> --dataset_avoided <sub_dataset_name> --saved_model <step1_save_path>/best_model.pth

Ensure that --dataset-avoided is set with the same value for both steps.

Inference

To evaluate the performance use the following commands:

python eval_step1.py --weight <step1_save_path>/best_model.pth --source_dataset_path <source_dataset> --target_dataset_path <target_dataset> --dataset_avoided <sub_dataset_name>

python eval_step2.py --weight <step2_save_path>/best_model.pth --source_dataset_path <source_dataset> --target_dataset_path <target_dataset> --dataset_avoided <sub_dataset_name>

Results

DA Method	Source (CrackSeg9k)	Target (BuildCrack)
AdaptSegnet*	47.53	48.47
MaxSquare	57.60	50.50
ADVENT*	47.51	48.47
IAST*	46.79	46.78
DAFormer*	47.54	48.47
DACS	58.46	58.11
CBST*	47.53	48.47
ProDA	50.32	47.94
FADA	79.18	60.73
CrackUDA	79.83	63.43

Citation

If you find the code and dataset useful, please cite this paper (and refer the dataset as BuildCrack):

@InProceedings{10.1007/978-3-031-78113-1_6,
author="Srivastava, Kushagra and Kancharla, Damodar Datta and Tahereen, Rizvi and Ramancharla, Pradeep Kumar
and Sarvadevabhatla, Ravi Kiran and Kandath, Harikumar",
title="CrackUDA: Incremental Unsupervised Domain Adaptation for Improved Crack Segmentation in Civil Structures",
booktitle="Pattern Recognition",
year="2025",
publisher="Springer Nature Switzerland",
address="Cham",
pages="74--89",
}

Acknowledgements

The authors acknowledge the financial support provided by IHuB-Data, IIIT Hyderabad to carry out this research work.

References

We thank the following github repositories for the relevant open source code and models:

• CrackSeg9k
• MDIL-SS