CarDD: A New Dataset for Vision-based Car Damage Detection

by Xinkuang Wang, Wenjing Li, Zhongcheng Wu.

CarDD Dataset

Information about the CarDD dataset is available at https://cardd-ustc.github.io/.

Car Damage Detection and Segmentation

Environment setup

Environment requirement: Pytorch 1.7.0 + Python 3.8 + CUDA 11.0

1. Clone the repo:

   git clone https://github.com/CarDD-USTC/CarDD-USTC.github.io.git
   cd CarDD-USTC.github.io/code/CarDD_detection
   

2. Prepare the environment:

   pip install openmim
   mim install mmdet
   pip install mmcv==1.7.0
   export MPLBACKEND='Agg' && export PYTHONPATH=$(CODE_PATH)/CarDD_detection/
   

Usage

1. Download CarDD at https://cardd-ustc.github.io/. Download pretrained models at Model Zoo to $(MODEL_PATH).

2. Train:

   python tools/train.py configs/car_damage/DCN_plus_cfg.py --work-dir $(WORK_PATH)
   

3. Test:

   python tools/test.py configs/car_damage/DCN_plus_cfg.py $(WORK_PATH)/epoch_24.pth --eval bbox segm --options "classwise=True"
   

4. Test and visualize:

   python tools/test.py configs/car_damage/DCN_plus_cfg.py $(WORK_PATH)/epoch_24.pth --show-dir $(VIS_PATH) --show-score-thr 0.7
   

5. Only inference:

   python tools/inference.py \
   --img-path=$(IMG_PATH) \
   --save-path=$(SAVE_PATH) \
   --config-file=configs/car_damage/DCN_plus_cfg.py  \
   --checkpoint-file=$(WORK_PATH)/epoch_24.pth
   

Salient Damage Detection

Environment setup

Please refer to each repository:

U2Net | PoolNet | KRN | CSNet | Saliency-Evaluation-Toolbox

Usage

• U2Net:

  cd $(CODE_PATH)/CarDD_SOD/U2-Net/
  train: python u2net_train.py
  test: python u2net_test.py
  

• PoolNet:

  cd $(CODE_PATH)/CarDD_SOD/PoolNet/
  train: python main.py --arch resnet --mode train --device 0 --data_root $(DATA_PATH)/CarDD_SOD/ --save_folder $(WORK_PATH)
  test: python main.py --mode test --model $(WORK_PATH)/run-0/models/final.pth --test_fold $(SAVE_PATH) --data_root $(DATA_PATH)
  

• KRN:

  cd $(CODE_PATH)/CarDD_SOD/KRN/
  train: python main_SGL_KRN.py --mode train --device 0 --data_root $(DATA_PATH)/CarDD_SOD/ --save_folder $(WORK_PATH)
  test: python main_SGL_KRN.py --mode test --device 0 --sal_mode t --test_model $(WORK_PATH)/run-0/models/final.pth --test_fold $(SAVE_PATH) --data_root $(DATA_PATH)
  

• CSNet:

  cd $(CODE_PATH)/CarDD_SOD/CSNet/CSNet_training/
  train: python train.py --config configs/csnet-L-x2_train-CarDD.yml
  test: python test.py --config configs/csnet-L-x2_train-CarDD.yml
  

• Evaluate:

  Please refer to Saliency-Evaluation-Toolbox.

Acknowledgments

• Segmentation code and models are from mmdetection.
• SOD code and models are respectively from U2Net | PoolNet | KRN | CSNet | Saliency-Evaluation-Toolbox.

Citation

If you found this code helpful, please consider citing:

@article{CarDD,
  author={Wang, Xinkuang and Li, Wenjing and Wu, Zhongcheng},
  journal={IEEE Transactions on Intelligent Transportation Systems}, 
  title={CarDD: A New Dataset for Vision-Based Car Damage Detection}, 
  year={2023},
  volume={24},
  number={7},
  pages={7202-7214},
  doi={10.1109/TITS.2023.3258480}
}