Implementation of a Crack Causal Augmentation Framework (CCAF) and Dynamic Binary Threshold (DBT) on Mask R-CNN for Crack Instance Segmentation

We provide the codes (Pytorch implementation), the datasets, and the pretrained model.

Pretrained Models

Pretrained models on PyTorch are available at Baidu Yun

passcode: oydm

Data

Dataset Format

This project uses the COCO dataset format and the data should be organized as follows

data_root/
├── images
│   ├── xxx.png
│   ├── xxy.png
│   └── ...
└── annotations
	├── train_coco_anns.json
	└── test_coco_anns.json

Dataset

CrackTunnel1K, Crack500, GAPs v2, CrackLS315 dataset are used in paper.

Our coco annotation files can download from Baidu Yun.

passcode：mm5i

Due to confidentiality, image data can be downloaded from the links provided in their corresponding papers.

CCAF

Using CCAF, the data is first enhanced by augmenting the train.json and test.json files in the annotations directory to train_imp.json and test_imp.json.

Then go training with command --imp True

python dilate_crack_coco.py --coco "path/to/train.json" --save "path/to/train_imp.json" --k 2
python dilate_crack_coco.py --coco "path/to/test.json" --save "path/to/test_imp.json" --k 2
python train.py --data-path "data_root_path" --imp True

Train DBT Mask R-CNN from scratch

Before you start training, please format the data to the above data format and set the dataloader correctly according to the script under the dataset dictionary.

Please download the resnet50 weights file and name it ./resnet50.pth.

• First train a Mask R-CNN model without DBT

python train.py --data-path "data_root_path" 

After training, select the appropriate weight file model_x.pth in the ./save_weights/ directory

• Second obtain the labels of DBT. You need to run the following two scripts to get the .csv file for the DBT

python obtain_dbt_label.py --data-path "data_root_path" --weights-path "model_x.pth" --save_pkl "othrs/labels_dbt.pkl"

python parse_best_othr.py --pkl_path "othrs/labels_dbt.pkl" --save_path "othrs/dbt_labels.csv"

• Final train Mask R-CNN with DBT

python train.py --data-path "data_root_path"  --DBT "True" --othrs "othrs/dbt_labels.csv"

The files in the directory ./save_weights are the Mask R-CNN with DBT weights after the training is completed.

Test

The test data format is still COCO format, please adapt it according to the corresponding file in the dataset directory

python test.py --data-path "data_root_path" --weights-path "weight-path" --dataset "CrackTunnel1K" 

We provide the test results of CrackTunnel1K.

Method	mAP50	MUCov	MWCov
Mask R-CNN	6.5	39.6	40
Mask R-CNN + CCAF + DBT	13	45.4	46.9

Predicted Results Visualization

Script predict.py provide predicted visualization for a single image

python predict.py --weight_path "weight_path" --img_path "img_path" 

The visualization results will be saved in the . /results directory

Some toy_images

Some toy_images can be found in dictionary ./toy_images

And the predicted results can be found in dictionary ./results

Image	Prediction

Copy Right

This dataset was collected for academic research.

Contact

For any problem about this dataset or codes, please contact Dr. Qin Lei (qinlei@cqu.edu.cn, tanlei086@gmail.com)