crackedcity/nvidia-hackathon-dataset

Crack & Pavement Segmentation — Curated Dataset Index

This repository is a curated reference index of public datasets relevant to crack and pavement segmentation. It does not redistribute any third-party images; it points to each dataset's canonical source so that researchers and practitioners can download them from the original authors under the original license terms.

For each dataset the index records: publisher, source link, brief domain description, approximate size, mask convention, license, and a short note.

Why this exists

Public crack-segmentation datasets are scattered across academic websites, Mendeley, Zenodo, GitHub, and HuggingFace. Conventions for the mask representation differ across datasets (some encode crack as 255, others as 0; some are single-band, others multi-class), which is a frequent source of silently broken training pipelines. This index consolidates the basic information needed to choose, download, and integrate each dataset, with the mask convention spelled out explicitly so it is harder to get wrong.

How to use

Programmatic access:

import json, urllib.request
url = "https://huggingface.co/datasets/crackedcity/nvidia-hackathon-dataset/resolve/main/catalog.json"
catalog = json.loads(urllib.request.urlopen(url).read())
for d in catalog["datasets"]:
    print(d["name"], "->", d["homepage"], "(", d["license"], ")")

Or read catalog.json directly.

Catalog

Dataset	Publisher	Domain	Approx. size	License	Source
CrackVision12K	Joo et al., 2025	Road pavement	12,000 (9,600 train / 1,200 val / 1,200 test)	see source	link
OmniCrack30K	Benz & Rodehorst, 2024	Multi-domain (road, masonry, concrete, asphalt)	~30,000 across train / val / test splits (a subset of files use a 1-pixel-skeleton annotation style and are typically filtered)	GPL-3.0	link
TACK Tunnel Data (TTD)	Sjölander et al., 2025	Tunnel linings (cracks, water, leaching)	3,774 images (785 with cracks, 197 water, 316 leaching)	see source	link
CrackStructures (and related close-up sets)	Benz et al., 2022	Close-up crack imagery on structural surfaces	Hundreds of high-resolution structural close-ups	see source	link
CRACK500	Yang et al., 2019	Smartphone pavement	500 source images, split into ~3,368 patches	see source	link
DeepCrack (Liu et al.)	Liu et al., 2019	Pavement and concrete	537 (300 train / 237 test)	see source	link
SegCODEBRIM	Mundt et al., 2019 (CODEBRIM) + segmentation re-annotation	Reinforced concrete bridge damage	Hundreds of close-up concrete shots with multi-label damage	see source	link

Dataset details

CrackVision12K

• Publisher / paper: Joo et al., 2025 — UCL Research Data Repository, 2024
• Domain: Road pavement
• Approx. size: 12,000 (9,600 train / 1,200 val / 1,200 test)
• Mask convention: crack = 255 (white) on black background
• License: see source
• Source / download: https://rdr.ucl.ac.uk/articles/dataset/CrackVision12K/26946472
• Note: Visit the source link for the canonical download and license terms.

OmniCrack30K

• Publisher / paper: Benz & Rodehorst, 2024 — CVPRW 2024
• Domain: Multi-domain (road, masonry, concrete, asphalt)
• Approx. size: ~30,000 across train / val / test splits (a subset of files use a 1-pixel-skeleton annotation style and are typically filtered)
• Mask convention: crack = 0 (black) on white background — inverse of most other datasets
• License: GPL-3.0
• Source / download: https://github.com/ben-z-original/omnicrack30k
• Note: Redistribution must inherit GPL-3.0 and include attribution. Visit the source.

TACK Tunnel Data (TTD)

• Publisher / paper: Sjölander et al., 2025 — arXiv:2512.14477
• Domain: Tunnel linings (cracks, water, leaching)
• Approx. size: 3,774 images (785 with cracks, 197 water, 316 leaching)
• Mask convention: Multi-class single-band PNG: 0=background, 40=crack, 160=water, 200=leaching
• License: see source
• Source / download: https://huggingface.co/datasets/CrackedCity/tack-tunnel-data
• Note: Visit the source for the canonical download and license terms.

CrackStructures (and related close-up sets)

• Publisher / paper: Benz et al., 2022
• Domain: Close-up crack imagery on structural surfaces
• Approx. size: Hundreds of high-resolution structural close-ups
• Mask convention: Binary masks; check source for exact convention
• License: see source
• Source / download: https://github.com/ben-z-original/crackstructures
• Note: Often used as a held-out, out-of-distribution evaluation set.

CRACK500

• Publisher / paper: Yang et al., 2019 — Feature Pyramid and Hierarchical Boosting Network for Pavement Crack Detection, TITS
• Domain: Smartphone pavement
• Approx. size: 500 source images, split into ~3,368 patches
• Mask convention: Binary, crack = 255
• License: see source
• Source / download: https://github.com/fyangneil/pavement-crack-detection
• Note: Widely cited baseline; visit the source for the canonical download.

DeepCrack (Liu et al.)

• Publisher / paper: Liu et al., 2019 — Neurocomputing
• Domain: Pavement and concrete
• Approx. size: 537 (300 train / 237 test)
• Mask convention: Binary masks, crack = 255
• License: see source
• Source / download: https://github.com/yhlleo/DeepCrack
• Note: Small but high-quality benchmark.

SegCODEBRIM

• Publisher / paper: Mundt et al., 2019 (CODEBRIM) + segmentation re-annotation — CVPR 2019
• Domain: Reinforced concrete bridge damage
• Approx. size: Hundreds of close-up concrete shots with multi-label damage
• Mask convention: Standard binary (crack = 255)
• License: see source
• Source / download: https://zenodo.org/record/2620293
• Note: Original CODEBRIM is classification; the seg re-annotation is community-driven.

Important notes on use

• Verify the license on the source page before redistributing. This index is not authoritative on license terms; it points to where the canonical license is published. Some datasets are GPL-3.0 (require source disclosure and license inheritance on derivative works), some are CC BY-NC (no commercial use), and some have custom academic-only terms.
• Mask convention is the most common gotcha. Always probe a handful of masks per source after extraction and confirm that the positive_frac (fraction of pixels above 127, or whatever your binarisation rule is) lies in a sensible range — for crack data that is usually 0.0001 to 0.05. A per-source value above 0.5 typically indicates an inverted convention.
• Citations. Each entry above lists the original authors and (where available) publication. Cite them when you use their data.

License (this index)

The curation, descriptions, and catalog.json in this repository are released under the Apache License 2.0. Each linked dataset is governed by its own license, listed on the source page.

Contributing

Found an out-of-date link, missed dataset, or license correction? Open an issue or PR on the repository.