Crack Detection as a Weakly-Supervised Problem: Towards Achieving Less Annotation-Intensive Crack Detectors

⚠️There are few bugs in commits earlier than Decenber 28, 2020. Please clone the newest commit to avoid having to debug them yourself.

Official repository of our ICPR2020 paper, "Crack Detection as a Weakly-Supervised Problem: Towards Achieving Less Annotation-Intensive Crack Detectors." arxiv

You will find the following in this repository:

• A script that downloads Aigle, CFD, and DeepCrack datasets
• URL for the low quality annotations repo (Rough, Rougher, Dil1-4 annotations used in the paper)
• Codes to generate your own synthetic annotations
• Setup script for the two crack detector OSS's used in the paper (DeepCrack and DeepLab v3)

Got no time to read? Here is a condensed README that lists the necessary commands to get you started.

Start the installation process with cloning this repo:

git clone --recursive https://github.com/hitachi-rd-cv/weakly-sup-crackdet.git

Requirements

Experiments were conducted on Ubuntu 18.04 with Python 3.6.9 and CUDA9. Other dependencies are summarized in requirements.txt. Be aware that GPUs with Turing architecture, such as RTX series are not compatible with CUDA9.

Data Preparation

Downloading the original annotations

In our experiments, the following datasets are used for training and testing:

• Aigle
• Crack Forest Dataset (CFD)
• DeepCrack Dataset (DCD)

These datasets are publicly available through different websites and GitHub repos. Run the following line to download the images and annotations

./tools/download.sh

Downloaded datasets should be available under data/*_detailed, where * stands for the name of the dataset.

• This script downloads images from different websites, and thus it may fail depending on the internet connections. Refer below for backup steps:
• Fails on the curl call to https://www.irit.fr/~Sylvie.Chambon/AigleRN_GT.html
  • Go to the website, copy the html and place it under data/aigle_github/tmp.html

Low Quality Annotation Repo

The proposed method was tested with various low quality annotations. Both manual and synthetic annotations are available through the Zenodo repo. After downloading the datasets, please locate them under data directory. Also note that the downloaded dataset only contains the annotations. Please run the following line to copy the RGB input images from the data/*_detailed directories.

python tools/data_gen.py --fill

⚠️ Make sure that all folders under data directory start with aigle_, cfd_, or deepcrack_

The Zenodo repo also contains pascal_voc_seg folder, which contains the pretrained Xception backbone for DeepLab. Place the folder under tools/model_supp/deeplab/datasets/pascal_voc_seg/.

Synthesizing your own dataset

You may want to synthesize your own low quality annotations. This can be done with the following line.

# generate synthetic dataset for the Aigle dataset, with dilation values 1, 2, 3, and 4
python tools/data_gen.py --synth_dil_anno --anno_type 1 2 3 4 --dataset_name aigle

--anno_type specifies the dilation value (n_{dil} in the paper). Larger value implies lower quality annotation. --dataset_name specifies the dataset name. Set all for synthesizing annotations for aigle, cfd, and deepcrack.

Formatting and copying the datasets

⚠️ This step needs to be done after the crack detectors are downloaded- i.e. after running the tools/setup_models.sh script

The datasets under data directory need to be formatted to be used by different crack detectors. This can be done with the following lines.

# format and dispatch detailed and dil1 annotations for used by DeepCrack
python tools/data_gen.py --deepcrack --anno_type detailed 1 --dataset_name all

# format and dispatch dil1 dil2 and rough annotations for used by DeepLab
python tools/data_gen.py --deeplab --anno_type 1 2 rough --dataset_name all

This script also copies the formatted annotations to the correct data directories within the downloaded repos. For DeepCrack repo, the data directory is ${DEEPCRACK_REPO}/datasets, and for DeepLab repo, it is ${DEEPLAB_REPO}/research/deeplab/datasets/data and ${DEEPLAB_REPO}/research/deeplab/datasets/data/tfrecords.

Setting up the Crack Detectors

Run the following script to download and modify the crack detector repos.

./tools/setup_models.sh --deepcrack --deeplab

This script should correctly set up the two crack detector repos under models directory. Please refer to the following sections for more details on what the script does.

⚠️ Do not forget to copy the datasets to the crack detector repos before training them. You can copy the dataset by following the instructions outlined in "Formatting and copying the datasets" section above.

DeepCrack

We use the 50440b52ddaf49cf54c2415e6b40646a7601c219 commit of the DeepCrack repo. ./tools/setup_models.sh clones and checks out the repo and modifies the repo by copying over files under tools/model_supp/deepcrack.

Training

You can train the DeepCrack model by running scripts/train_deepcrack.sh from models/deepcrack directory. Modify the script accordingly to train the model with various annotations. The training results are saved under the checkpoints directory.

Evaluation

You can evaluate the DeepCrack model by running scripts/test_eval.sh from models/deepcrack directory. Modify the script accordingly to evaluate the model outputs for various annotations. The training results are saved under the checkpoints directory.

For more details on training and evaluation, please refer to the original repository.

DeepLab

We use the 0a161121852ee5f34b939279d54b5d3e231ca501 commit of the DeepLab repo (sorry it is an old commit, the recent repository uses TF v2 instead of v1). ./tools/setup_models.sh clones and checks out the repo and modifies the repo by copying over files under tools/model_supp/deeplab.

Training

Before training the model, this repo requires that $PYTHONPATH environment variable is properly set. Run the following lines:

cd models/deeplab/research
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

You can train the DeepLab model by running scripts/train.sh from models/deeplab/research/deeplab directory. Modify the script accordingly to train the model with various annotations. The training results are saved under the outputs directory.

Evaluation

You can evaluate the DeepLab model by running scripts/inference.sh from models/deeplab/research/deeplab directory. Modify the script accordingly to evaluate the model outputs for various annotations. The training results are saved under the outputs directory.

For more details on training and evaluation, please refer to the original repository.

Inoue et. al.

Unfortunately, we cannot release this code due to company confidentiality reasons.

Evaluating with the Micro Branch

To evaluate the trained models with the Micro Branch, copy the evaluation results from the crack detector repos to the eval/results directory. Results are stored under models/deepcrack/checkpoints for DeepCrack, and results are stored under models/deeplab/research/deeplab/outputs for DeepLab.

Evaluation can be done as follows:

python eval/micro_eval.py

This script first applies the Micro Branch output to the crack detector outputs and stores the results under eval/results/${MODEL_RESULT_FOLDERNAME}/cv_output_none. Then the script evaluates the result and outputs eval_cv_output_none_dil0.txt and eval_test_output_dil0.txt, which correspond to results with Micro Branch and without Micro Branch, respectively.

The result for all model results can be aggregated using the gen_table.py script. It outputs a csv file named results.csv under the main directory.

Citation

Please consider citing our paper if it helps your research:

@inproceedings{inoue2020crack,
    title={Crack Detection as a Weakly-Supervised Problem: Towards Achieving Less Annotation-Intensive Crack Detectors},
    author={Inoue, Yuki and Nagayoshi, Hiroto},
    booktitle={International Conference on Pattern Recognition (ICPR)},
    year={2020},
}

Acknowledgement

Our project is built from the following repositories. Thanks you for your great works!

• Models
  • DeepCrack
  • DeepLab
• Datasets
  • Aigle
  • Crack Forest Dataset
  • DeepCrack