OpenCrack

A consolidated, leakage-controlled benchmark for pavement-crack segmentation, in COCO format.

OpenCrack unifies dozens of separately-annotated public crack datasets into a single benchmark with a shared four-class taxonomy, cross-split deduplication, and a fixed stratified split. It is released in two regimes:

• OpenCrack: the in-domain, close-up crack imagery (asphalt, concrete, masonry, facades, bridge decks). This is the primary pixel-mask benchmark.
• PaveSafe: an out-of-domain deployment benchmark drawn from operational road-inspection systems with wider acquisition geometry, retaining non-crack distress classes (potholes, patches, manholes).

The two-regime design separates in-domain accuracy from cross-regime transfer, and the consolidation removes the train/test image leakage that inflates scores on the large composite crack datasets.

This repository is the benchmark's public reference page: the dataset roster, the model-comparison results, and full source citations. See Citing this work to reference the benchmark.

---

At a glance

	Images (dedup)	Annotations (dedup)	Images (full)	Annotations (full)
OpenCrack (in-domain)	82,218	99,493	104,930	129,116
PaveSafe (out-of-domain)	48,770	105,583	50,522	110,987

• 38 public source datasets consolidated: 32 into OpenCrack (under 28 COCO dataset_source labels, six nested in composite archives) and 6 into PaveSafe.
• Mixed annotation granularity: pixel masks, bounding boxes, and confirmed-negative (crack-free) sources, unified in a COCO instance format. Segmentation metrics are computed on the pixel-mask subset.
• Confirmed-negative pool of 11,229 crack-free images (8,034 from SDNET2018 uncracked splits + 3,195 from the Bridge Crack Library noncrack set), reserved for false-positive-rate evaluation.
• Deduplication with DINOv2 image embeddings at a conservative cosine threshold (τ = 0.979), so no near-duplicate image straddles the train/validation/test boundary.
• Stratified 70/15/15 split on source dataset, crack presence, and crack-width band.

---

Four-class taxonomy

The taxonomy is anchored to engineering inspection standards (ASTM/PCI, GOST), so the labels map to the distress families that road authorities track and repair.

ID	Class	Definition
1	material_discontinuity	Cracks: surface fractures from exceeded tensile strength (longitudinal, transverse, diagonal, alligator, block, map-cracking), and sealed/repaired cracks
2	surface_openings	Structural material loss: potholes, spalling, voids where the surface is missing
3	patching_and_repairs	Prior maintenance interventions: surface and full-depth patches
4	functional_infrastructure	Non-distress structural components that generate false positives: manhole covers, drainage grates

Sealed and repaired cracks are merged into class 1; the source label is preserved in an original_class field on every annotation. OpenCrack populates mainly classes 1–2; classes 3–4 are populated by PaveSafe's multi-class sources.

---

Results: model comparison

Four models are evaluated on exactly the same held-out OpenCrack stratified test cohort (6,910 crack-pixel-annotated positives), with one harness and 1,000-resample bootstrap confidence intervals. The compared models are three released baselines (CrackSAM-adapter [11], an adapter on Segment Anything; OmniCrack30k [4], an nnU-Net trained on a prior composite; and Hybrid-Segmentor [12], a CNN-transformer) and an nnU-Net [14] trained on the OpenCrack training split.

In-domain (OpenCrack)

IoU and clIoU(τ=4) as image-weighted (micro) means over the 6,910 positives, with the per-source (macro) mean alongside:

Model	IoU micro [95% CI]	IoU macro	clIoU(τ=4) micro [95% CI]	clIoU macro
nnU-Net (trained on OpenCrack)	0.539 [0.534, 0.545]	0.470	0.641 [0.634, 0.648]	0.691
CrackSAM-adapter [11]	0.470 [0.463, 0.476]	0.391	0.606 [0.598, 0.613]	0.669
OmniCrack30k [4]	0.426 [0.420, 0.433]	0.384	0.557 [0.548, 0.565]	0.633
Hybrid-Segmentor [12]	0.402 [0.395, 0.409]	0.265	0.468 [0.460, 0.476]	0.379

The trained nnU-Net leads on both metrics with confidence-interval-disjoint margins. The comparison against OmniCrack30k is the informative one: both are the same nnU-Net v2 self-configuring pipeline, applied to a different training corpus (a prior composite instead of OpenCrack). Our model leads by +0.11 IoU (and +0.058 clIoU macro, 0.691 vs 0.633) with a smaller training budget — a single fold at 500 epochs, against OmniCrack30k's default schedule (a five-fold ensemble at nnU-Net's default 1,000 epochs). The gap therefore reflects the value of the OpenCrack training data within the same automated method, not architecture or budget tuning.

Full five-metric matrix (OpenCrack, n = 6,910)

The ranking is identical across all five per-image metrics, so no conclusion depends on the metric choice.

Model	IoU	Dice	Boundary F1	NSD	clIoU(τ=4)
nnU-Net (this work)	0.539 [0.534, 0.545]	0.664 [0.658, 0.669]	0.608 [0.601, 0.615]	0.700 [0.694, 0.707]	0.641 [0.634, 0.648]
CrackSAM-adapter [11]	0.470 [0.463, 0.476]	0.586 [0.578, 0.592]	0.523 [0.516, 0.531]	0.631 [0.623, 0.638]	0.606 [0.598, 0.613]
OmniCrack30k [4]	0.426 [0.420, 0.433]	0.532 [0.524, 0.539]	0.504 [0.496, 0.513]	0.587 [0.578, 0.597]	0.557 [0.548, 0.565]
Hybrid-Segmentor [12]	0.402 [0.395, 0.409]	0.493 [0.485, 0.501]	0.463 [0.456, 0.471]	0.537 [0.529, 0.545]	0.468 [0.460, 0.476]

Cross-regime transfer (PaveSafe, n = 730)

Models trained or fine-tuned on close-up crack imagery do not transfer to PaveSafe's wider road scenes: the in-domain leader falls from 0.539 to 0.142 IoU, and every crack model collapses into a narrow 0.09–0.14 band. The ranking is regime-dependent, not a fixed ordering of model quality.

Model	IoU	Dice	Boundary F1	NSD	clIoU(τ=4)
nnU-Net (this work)	0.142 [0.128, 0.155]	0.203 [0.184, 0.221]	0.206 [0.189, 0.223]	0.257 [0.236, 0.278]	0.201 [0.182, 0.221]
OmniCrack30k [4]	0.124 [0.110, 0.138]	0.176 [0.157, 0.193]	0.209 [0.189, 0.230]	0.241 [0.219, 0.264]	0.192 [0.173, 0.212]
CrackSAM-adapter [11]	0.097 [0.084, 0.110]	0.137 [0.120, 0.152]	0.050 [0.043, 0.058]	0.077 [0.066, 0.089]	0.132 [0.117, 0.147]
Hybrid-Segmentor [12]	0.090 [0.077, 0.104]	0.126 [0.109, 0.143]	0.048 [0.041, 0.054]	0.072 [0.062, 0.081]	0.053 [0.045, 0.061]

Per-source breakdowns (by crack width and surface material), the hard-negative enrichment study, and the reproduction of published baseline numbers are part of the full evaluation accompanying this benchmark.

---

The trained model

The supervised baseline is the self-configuring nnU-Net v2 (2D configuration) [14], trained from random initialisation on the OpenCrack class-1 training split:

• Plain convolutional U-Net selected by the nnU-Net planner (seven stages, feature widths [32, 64, 128, 256, 512, 512, 512]), InstanceNorm + LeakyReLU, ≈92.5 M parameters (counted from the released checkpoint).
• 256×256 patches, per-image z-normalisation, nnU-Net default augmentation; SGD default schedule.
• A fixed training budget of about 500 epochs, single fold (fold 0), no early stopping and no multi-fold ensembling, matched to the single-fold OmniCrack30k so the head-to-head isolates training data from architecture.

Weights: the trained checkpoint and planner-generated configuration are on Hugging Face: fadeevla/opencrack-nnunet.

---

Data and reproduction

The source datasets are not redistributed in this repository: each retains its own license and must be obtained from its original provider (see References). What is published here is the consolidated benchmark's specification, roster, and results.

The consolidation pipeline (taxonomy mapping, mask-to-RLE conversion, DINOv2 deduplication, stratified splitting) and the unified evaluation harness are maintained in the project's development repository.

---

OpenCrack source datasets

32 source datasets under 28 COCO dataset_source labels (six are nested inside composite archives). Image counts are for the deduplicated build. All sources provide a pixel mask per image except CCIC and SDNET2018, which are image-level crack/non-crack sources and supply most of the confirmed-negative pool.

Source	Images	Source	Images
CCIC [22]	17,763	DeepCrack [20]	365
SDNET2018 [9]	17,170	Facade390 †	364
CrackVision12K [12]	9,484	BuildCrack [26]	354
Conglomerate [5]	7,834	Masonry [8]	239
Bridge Crack Library [16]	7,745	CrackLS315 [20]	194
TopoDS [23]	7,047	CrackSC †	168
CRACK500 [30]	3,280	CrackNJ156 [29]	155
Concrete3k [19]	2,891	CrackTree260 [31]	150
Asphalt3k [19]	2,273	Ceramic [17]	100
UDTIRI-Crack [13]	2,229	AEL [1]	57
DIC [25]	530	CRKWH100 [20]	44
S2DS [3]	497	Khanh11k †	41
SegCODEBRIM [15]	441	CrackSeg9k [18]	7
Road420 †	415	GAPS384 [10]	381

† Dataset release with no associated publication; distributed as data only (Road420, Facade390, CrackSC, Khanh11k).

PaveSafe source datasets

Dataset	Images	Annotation	Surface	Geography
RDD2022 [2]	28,000	Bounding box (4 classes)	Asphalt road	6 countries
SVRDD [24]	7,999	Bounding box (7 classes)	Asphalt road	China
Mosquitonet / MAPSIA [6]	6,424	Bounding box (13 classes)	Road	Spain
PavementScapes [27]	3,305	Pixel mask	Asphalt (top-down)	China
CrackSeg [28]	2,366	Pixel mask	Asphalt (oblique)	Mixed
UDTIRI pothole [13]	676	Pixel mask (pothole)	Road	China

---

Citing this work

If you use OpenCrack, please cite this work and credit each source dataset you rely on (the references below).

@misc{fadeev2026opencrack,
  author       = {Fadeev, V. A.},
  title        = {OpenCrack: A Consolidated, Leakage-Controlled Benchmark for Pavement Crack Segmentation},
  year         = {2026},
  howpublished = {\url{https://github.com/fadeevla/OpenCrack}}
}

---

License and attribution

The contents of this repository (the README, the roster, the compiled results, and the presentation) are released under CC-BY-4.0. This license covers the benchmark specification and documentation only; it does not relicense the underlying images or annotations. Every source dataset remains under its own original license; obtain each from its provider and comply with its terms.

---

References

1. Amhaz, R. Automatic Crack Detection on Two-Dimensional Pavement Images: An Algorithm Based on Minimal Path Selection / R. Amhaz, S. Chambon, J. Idier, V. Baltazart // IEEE Transactions on Intelligent Transportation Systems. — 2016. — Vol. 17, no. 10. — P. 2718–2729. — DOI 10.1109/TITS.2015.2477675.
2. Arya, D. RDD2022: A multi-national image dataset for automatic road damage detection / D. Arya, H. Maeda, S. K. Ghosh, D. Toshniwal, Y. Sekimoto // Geoscience Data Journal. — 2024. — DOI 10.1002/gdj3.260.
3. Benz, C. Image-Based Detection of Structural Defects Using Hierarchical Multi-scale Attention / C. Benz, V. Rodehorst // Pattern Recognition: 44th DAGM German Conference (DAGM GCPR 2022). — Springer, 2022. — P. 337–353. — DOI 10.1007/978-3-031-16788-1_21.
4. Benz, C. OmniCrack30k: A Benchmark for Crack Segmentation and the Reasonable Effectiveness of Transfer Learning / C. Benz, V. Rodehorst // 2024 IEEE/CVF CVPR Workshops (CVPRW). — IEEE, 2024. — P. 3876–3886. — DOI 10.1109/CVPRW63382.2024.00392.
5. Bianchi, E. Concrete Crack Conglomerate Dataset / E. Bianchi, M. Hebdon. — 2021. — DOI 10.7294/16625056.
6. Cano-Ortiz, S. An end-to-end computer vision system based on deep learning for pavement distress detection and quantification / S. Cano-Ortiz, L. Lloret Iglesias, P. Martinez Ruiz Del Árbol, P. Lastra-González, D. Castro-Fresno // Construction and Building Materials. — 2024. — Vol. 416. — Art. 135036. — DOI 10.1016/j.conbuildmat.2024.135036.
7. Carion, N. SAM 3: Segment Anything with Concepts / N. Carion [et al.]. — Meta AI, 2025.
8. Dais, D. Automatic crack classification and segmentation on masonry surfaces using convolutional neural networks and transfer learning / D. Dais, İ. E. Bal, E. Smyrou, V. Sarhosis // Automation in Construction. — 2021. — Vol. 125. — Art. 103606. — DOI 10.1016/j.autcon.2021.103606.
9. Dorafshan, S. SDNET2018: An annotated image dataset for non-contact concrete crack detection using deep convolutional neural networks / S. Dorafshan, R. J. Thomas, M. Maguire // Data in Brief. — 2018. — Vol. 21. — P. 1664–1668. — DOI 10.1016/j.dib.2018.11.015.
10. Eisenbach, M. How to get pavement distress detection ready for deep learning? A systematic approach / M. Eisenbach, R. Stricker, D. Seichter [et al.] // 2017 International Joint Conference on Neural Networks (IJCNN). — IEEE, 2017. — P. 2039–2047. — DOI 10.1109/IJCNN.2017.7966101.
11. Ge, K. Fine-tuning vision foundation model for crack segmentation in civil infrastructures / K. Ge, C. Wang, Y. Guo, Y. Tang, Z. Hu, H. Chen // arXiv. — 2024. — arXiv:2312.04233.
12. Goo, J. M. Hybrid-Segmentor: A Hybrid Approach to Automated Fine-Grained Crack Segmentation in Civil Infrastructure / J. M. Goo, X. Milidonis, A. Artusi, J. Boehm, C. Ciliberto // Automation in Construction. — 2025. — Vol. 170. — Art. 105960. — DOI 10.1016/j.autcon.2024.105960.
13. Guo, S. UDTIRI: An Online Open-Source Intelligent Road Inspection Benchmark Suite / S. Guo, J. Li, Y. Feng [et al.] // IEEE Transactions on Intelligent Transportation Systems. — 2024. — Vol. 25, no. 8. — P. 9920–9931. — DOI 10.1109/TITS.2024.3351209.
14. Isensee, F. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation / F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, K. H. Maier-Hein // Nature Methods. — 2021. — Vol. 18, no. 2. — P. 203–211. — DOI 10.1038/s41592-020-01008-z.
15. Jaziri, A. SegCODEBRIM / A. Jaziri, M. Mundt, A. F. Rodriguez, V. Ramesh. — 2023. — DOI 10.5281/zenodo.10071534.
16. Jin, T. Bridge Crack Library / T. Jin, Z. Li, Y. Ding, S. Ma, Y. Ou. — 2021. — DOI 10.7910/DVN/RURXSH.
17. Junior, G. Ceramic Cracks Segmentation with Deep Learning / G. Junior, J. Ferreira, C. Millán-Arias, R. Daniel, A. Casado, B. Fernandes // Applied Sciences. — 2021. — Vol. 11. — Art. 6017. — DOI 10.3390/app11136017.
18. Kulkarni, S. CrackSeg9k: A Collection and Benchmark for Crack Segmentation Datasets and Frameworks / S. Kulkarni, S. Singh, D. Balakrishnan, S. Sharma, S. Devunuri, S. C. R. Korlapati // arXiv. — 2022. — arXiv:2208.13054.
19. Li, Y. Real-time High-Resolution Neural Network with Semantic Guidance for Crack Segmentation / Y. Li, R. Ma, H. Liu, G. Cheng // Automation in Construction. — 2023. — Vol. 156. — Art. 105112. — DOI 10.1016/j.autcon.2023.105112.
20. Liu, Y. DeepCrack: A deep hierarchical feature learning architecture for crack segmentation / Y. Liu, J. Yao, X. Lu, R. Xie, L. Li // Neurocomputing. — 2019. — Vol. 338. — P. 139–153. — DOI 10.1016/j.neucom.2019.01.036.
21. Oquab, M. DINOv2: Learning Robust Visual Features without Supervision / M. Oquab, T. Darcet, T. Moutakanni [et al.] // arXiv. — 2023. — arXiv:2304.07193.
22. Özgenel, Ç. F. Performance Comparison of Pretrained Convolutional Neural Networks on Crack Detection in Buildings / Ç. F. Özgenel, A. G. Sorguç // 35th International Symposium on Automation and Robotics in Construction (ISARC). — 2018. — DOI 10.22260/ISARC2018/0094.
23. Pantoja-Rosero, B. G. TOPO-Loss for continuity-preserving crack detection using deep learning / B. G. Pantoja-Rosero, D. Oner, M. Kozinski [et al.] // Construction and Building Materials. — 2022. — Vol. 344. — Art. 128264. — DOI 10.1016/j.conbuildmat.2022.128264.
24. Ren, M. An annotated street view image dataset for automated road damage detection / M. Ren, X. Zhang, X. Zhi, Y. Wei, Z. Feng // Scientific Data. — 2024. — Vol. 11, no. 1. — Art. 407. — DOI 10.1038/s41597-024-03263-7.
25. Rezaie, A. Comparison of Crack Segmentation Using Digital Image Correlation Measurements and Deep Learning / A. Rezaie, R. Achanta, M. Godio, K. Beyer // Construction and Building Materials. — 2020. — Vol. 261. — Art. 120474. — DOI 10.1016/j.conbuildmat.2020.120474.
26. Srivastava, K. CrackUDA: Incremental Unsupervised Domain Adaptation for Improved Crack Segmentation in Civil Structures / K. Srivastava, D. Kancharla, R. Tahereen, P. Ramancharla, S. R. K. Sarvadevabhatla, H. Kandath // arXiv. — 2024. — arXiv:2412.15637.
27. Tong, Z. Pavementscapes: a large-scale hierarchical image dataset for asphalt pavement damage segmentation / Z. Tong, T. Ma, J. Huyan, W. Zhang // arXiv. — 2022. — arXiv:2208.00775.
28. TRMetaGroup. CrackSeg : dataset. — 2025. — URL: https://github.com/TRMetaGroup/CrackSeg.
29. Xu, Z. Pavement crack detection from CCD images with a locally enhanced transformer network / Z. Xu, H. Guan, J. Kang [et al.] // International Journal of Applied Earth Observation and Geoinformation. — 2022. — Vol. 110. — Art. 102825. — DOI 10.1016/j.jag.2022.102825.
30. Yang, F. Feature Pyramid and Hierarchical Boosting Network for Pavement Crack Detection / F. Yang, L. Zhang, S. Yu, D. Prokhorov, X. Mei, H. Ling // IEEE Transactions on Intelligent Transportation Systems. — 2020. — Vol. 21, no. 4. — P. 1525–1535. — DOI 10.1109/TITS.2019.2910595.
31. Zou, Q. CrackTree: Automatic crack detection from pavement images / Q. Zou, Y. Cao, Q. Li, Q. Mao, S. Wang // Pattern Recognition Letters. — 2012. — Vol. 33, no. 3. — P. 227–238. — DOI 10.1016/j.patrec.2011.11.004.