3DCrack: A 3D Laser-Scanned Pavement Crack Segmentation Dataset

DOI: 10.21227/2etq-xh36 · Paper: arXiv:2508.10256 · Code: Awesome-Crack-Detection · Mirrors: IEEE DataPort, Zenodo

3DCrack is a pixel-level pavement crack segmentation dataset built from 3D laser scans, with matching crack centerline (vector) annotations. It was introduced in the review paper "Deep Learning for Crack Detection: A Review of Learning Paradigms, Generalizability, and Datasets" (arXiv:2508.10256), where it is used to benchmark common deep learning methods, including recent foundation models.

Data source

The images are range (depth) images of pavement surfaces, captured with a 3D line laser imaging system used for pavement surveys. The data was collected across the state of Georgia, USA, over 10 years by the Smart City Infrastructure (SCI) Group at the Georgia Institute of Technology. Each image is 512 × 624 pixels (width × height), at about 4 mm per pixel.

Contents

Split	Images	Masks	Centerline (CVM) files	Crack-free
Train	1,139	1,139	870	269
Val	245	245	185	60
Test	250	250	197	53
Total	1,634	1,634	1,252	382

The 382 samples without a CVM file are crack-free images; their masks are all background.

Quick start

from datasets import load_dataset
import json

ds = load_dataset("THiNK327/3DCrack")  # splits: train / validation / test
s = ds["train"][0]
s["image"], s["mask"]                  # PIL images (512 × 624)
cvm = json.loads(s["cvm_geojson"]) if s["cvm_geojson"] else None

Fields

Field	Type	Description
id	string	Sample stem, e.g. sample_000002
image	image	Range image (8-bit PNG)
mask	image	Binary crack mask (0 = background, 255 = crack)
has_crack	bool	False for crack-free samples (no CVM file)
cvm_geojson	string | null	Raw CVM GeoJSON text (see below); null for crack-free samples

Original file layout

The original archive is also provided unchanged as raw/3DCrack_v1.zip (md5 62188344ebdb4a4fef2618de9471df50, identical to the IEEE DataPort and Zenodo releases):

3DCrack_v1/
├── Train/ | Val/ | Test/
│   ├── img/   sample_XXXXXX.png       range image (512 × 624)
│   ├── anno/  sample_XXXXXX.png       binary crack mask (0 = background, 255 = crack)
│   └── CVM/   sample_XXXXXX.geojson   crack centerlines with per-point widths
├── histograms/per_split_panels.png    statistics for each split
└── vis_cvm.py                         viewer: image, mask and centerlines side by side

Files with the same stem (for example sample_000002) belong to the same sample. To browse a split with the viewer: python vis_cvm.py Test.

CVM centerline format

Each .geojson file is a FeatureCollection with one LineString feature per crack branch. Coordinates are pixel positions stored as [row, column], not [x, y]. Each feature's properties hold two per-vertex crack widths: width_mm, measured from the crack map, and width_mm_rangeimg, measured from the range image. A value of 0 is a placeholder at branch endpoints. The top-level properties.image gives the image size.

License

Creative Commons Attribution 4.0 International (CC BY 4.0). Copyright (C) 2026 The Authors.

Citation

If you use this dataset, please cite both the dataset (DOI 10.21227/2etq-xh36) and the paper:

@misc{zhang2025_3dcrack,
  title     = {3DCrack: A 3D Laser-Scanned Pavement Crack Segmentation Dataset},
  author    = {Zhang, Xinan and Wang, Haolin and Hsieh, Yung-An and Yang, Zhongyu and Yezzi, Anthony and Tsai, Yi-Chang},
  year      = {2025},
  publisher = {IEEE DataPort},
  doi       = {10.21227/2etq-xh36},
  url       = {https://doi.org/10.21227/2etq-xh36}
}

@article{zhang2025crackreview,
  title   = {Deep Learning for Crack Detection: A Review of Learning Paradigms, Generalizability, and Datasets},
  author  = {Zhang, Xinan and Wang, Haolin and Hsieh, Yung-An and Yang, Zhongyu and Yezzi, Anthony and Tsai, Yi-Chang},
  journal = {arXiv preprint arXiv:2508.10256},
  year    = {2025}
}

Authors

Xinan Zhang, Haolin Wang, Yung-An Hsieh, Zhongyu Yang, Anthony Yezzi, Yi-Chang Tsai — Georgia Institute of Technology.