This repository mainly includes 2 datasets. If you have any questions, please don't hesitate to contact us.
- 🍎 Steelcrack Dataset.
- 🍇 Crack dataset.
- 2024.02.29: We release Steelcrack dataset!
Download address: Google Drive or OneDrive.
All the images in the Steelcrack are directly captured from different projects of steel structures.
| Dataset | Image size | Training set | Validation set | Test set |
|---|---|---|---|---|
| Steelcrack | 512 × 512 | 3300 images | 525 images | 530 images |
Some of the images are from the 1st International Project Competition for SHM, while others are provided by us.
We re-label all the images to get more refined annotations.
To ensure comparability and reproducibility, all the models are trained from scratch using the same training policy.
| Method | mi IoU (%) | mi Dice (%) | #Param. (M) | MACs (G) |
|---|---|---|---|---|
| U-Net | 68.49 | 75.13 | 7.77 | 55.01 |
| U-Net++ | 72.23 | 78.37 | 9.16 | 138.63 |
| Attention U-Net | 71.25 | 77.54 | 34.88 | 266.54 |
| CE-Net | 76.00 | 81.54 | 29.00 | 35.60 |
| DeepLabv3+ (MobileNetv2) | 68.22 | 71.07 | 5.81 | 29.13 |
| DeepLabv3+ (Xception) | 67.40 | 71.48 | 54.70 | 83.14 |
| DeepLabv3+ (ResNet-101) | 69.04 | 69.45 | 59.34 | 88.84 |
| SCRN | 73.23 | 78.91 | 25.23 | 31.92 |
| TransUNet | 64.34 | 72.55 | 67.87 | 129.96 |
| CrackSeU-B | 70.42 | 80.50 | 3.19 | 11.22 |
| CrackSeU-L | 71.66 | 81.24 | 4.62 | 28.22 |
| DconnNet | 74.73 | 83.40 | 28.38 | 24.79 |
| BGCrack V1 | 77.16 | 85.33 | 2.32 | 15.76 |
You are very welcome to use and cite our datasets! The BibTeX entry is as follows:
@article{HE2024BGCrack,
title = {Crack segmentation on steel structures using boundary guidance model},
journal = {Automation in Construction},
volume = {162},
pages = {105354},
year = {2024},
issn = {0926-5805},
doi = {https://doi.org/10.1016/j.autcon.2024.105354},
url = {https://www.sciencedirect.com/science/article/pii/S0926580524000906},
author = {Zhili He and Wang Chen and Jian Zhang and Yu-Hsing Wang},
keywords = {Crack inspection, Deep learning, Boundary guidance method, Benchmark dataset}
}