tldr.md

Tl;dr

If README is too long to read, here is the stripped down version that just lists the commands to execute to train DeepCrack and DeepLab models with Dilation-1 Annotation for Aigle, CFD, and DCD.

Preliminaries

1. Clone the repo

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

2. Download the RGB images and the Precise Annotations

   ./tools/download.sh

   • 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

3. Download the low quality annotations from Zenodo repo, and copy them to the "data" directory

   cp -r ${weakly_sup_crackdet_dataset}/data/* ${git_repo_dir}/data

4. Download the pretrained Xception backbone network from the Zenodo repo and copy it to the main directory.

   cp -r ${pascal_voc_seg}/ tools/model_supp/deeplab/datasets/pascal_voc_seg/

5. Copy the RGB images from the data/*_detailed directories to the low quality annotations

   python tools/data_gen.py --fill

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

6. Set up the repos for the crack detectors

   ./tools/setup_models.sh

Training DeepCrack

1. Prepare the dataset

   python tools/data_gen.py --deepcrack --anno_type detailed rough rougher 1 2 3 4 --dataset_name all

2. Move to the DeepCrack directory

   cd models/deepcrack

3. Train the model. The script will train the model with synthetic annotation, dilation=1. Change the contents of the script accordingly to train with other annotations. The script also runs scripts/test_eval.sh at the end to format the training output.

   scripts/train_deepcrack.sh

   • ⚠️ The --name option for train.py specifies the output name, and it should be formatted as follows: ${DATASET_NAME}_deepcrack_${ANNOTATION_TYPE}.
   • After running scripts/test_eval.sh, you should be able to see the outputs under checkpoints/${DATASET_NAME}_deepcrack_${ANNOTATION_TYPE}/sample_imgs/test_output

Training DeepLab V3+

1. Prepare the dataset

   python tools/data_gen.py --deeplab --anno_type detailed rough rougher 1 2 3 4 --dataset_name all

2. Move to the research directory of the DeepLab directory and set the environmental variable as per requested by the DeepLab repo.

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

3. Move to the DeepLab directory

   cd deeplab

4. Train the model. The script will train the model with synthetic annotation, dilation=1. Change the contents of the script accordingly to train with other annotations. The script also runs scripts/inference.sh at the end to format the training output.

   scripts/train.sh

   • ⚠️ The --train_logdir option for train.py specifies the output name, and it should be formatted as follows: ${DATASET_NAME}/deeplab/${ANNOTATION_TYPE}.

Evaluation

1. Move the trained results from the crack detection repos to the main evaluation directory.

   # move to the main directory
   cd ${git_repo_dir}
   
   # create the eval/results directory
   mkdir eval/results
   
   # for DeepCrack
   mv models/deepcrack/checkpoints/* eval/results/
   
   # for DeepLab
   mv models/deeplab/research/deeplab/outputs/* eval/results/

2. Evaluate

   python eval/micro_eval.py

3. Aggregate the results. It outputs a csv file named results.csv

   python eval/gen_table.py

   • Contents of results.csv
     • f1_0_test: F-score of the output without the Micro Branch is applied
     • f1_0_cv_none: F-score of the output with the Micro Branch is applied