README.md

---
license: other
license_name: mixed-source-attribution
pretty_name: Unified Road Defect Dataset
task_categories:
  - object-detection
tags:
  - road-damage
  - pavement-distress
  - pothole-detection
  - crack-detection
  - yolo
  - autonomous-driving
  - infrastructure
size_categories:
  - 10K<n<100K
---

# Unified Road Defect Dataset

A merged, YOLO-format road-defect detection dataset that combines **RDD-2022**
(primary, ground-level, 6 countries) with two supplementary aerial/drone
datasets — **UAV-PDD2023** (China) and **RoadDamageVision** (China + Spain) —
into a single **4-class CRDDC** schema.

> **This is a derived dataset.** It re-packages and re-labels images from three
> independently published sources. All credit for the underlying images and
> original annotations belongs to their respective authors (see
> [Licensing & Attribution](#licensing--attribution)). If you use this dataset
> you **must cite all three source datasets** below.

## Classes (4-class CRDDC)

| id | code | name |
|----|------|------|
| 0 | D00 | Longitudinal Crack |
| 1 | D10 | Transverse Crack |
| 2 | D20 | Alligator Crack |
| 3 | D40 | Pothole |

## Layout (Ultralytics YOLO)

```
data/
  images/train/  images/val/
  labels/train/  labels/val/   # <class> <cx> <cy> <w> <h>  (normalized)
rdd_merged.yaml                 # Ultralytics data config (nc=4)
STATS.md                        # per-class / per-source / per-country counts
```

Every file is **source-prefixed** (`rdd_…`, `uav_…`, `rdv_…`) so per-source
metrics can be reported.

## How the sources were merged

| Source class | → unified | rationale |
|---|---|---|
| RDD-2022 D00/D10/D20/D40 | D00/D10/D20/D40 | primary schema |
| RDD-2022 other (D11/D43/D44/D50…) | dropped | not in CRDDC 4-class |
| UAV-PDD2023 LC/TC/AC/PH | D00/D10/D20/D40 | direct map |
| UAV-PDD2023 OC (oblique) / RP (repair) | dropped | no clean CRDDC equivalent |
| RoadDamageVision D00/D10/D20/D40 | D00/D10/D20/D40 | direct map (same D-codes) |
| RoadDamageVision Repair / Block Crack | dropped | no CRDDC 4-class equivalent |

---

## Licensing & Attribution

This derived dataset is released for **research purposes**. The three sources
were published under their own terms — **verify and comply with each on its
original host** before redistribution or commercial use. The combined release
inherits the most restrictive of the source terms.

### 1. RDD-2022 — *primary*
Multi-national Road Damage Dataset (CRDDC'2022), 47,420 images, 6 countries.
- Source: figshare `21431547` · sekilab/RoadDamageDetector
- License (per host): Creative Commons — verify on figshare before reuse.

```bibtex
@article{arya2022rdd2022,
  title   = {RDD2022: A multi-national image dataset for automatic road damage detection},
  author  = {Arya, Deeksha and Maeda, Hiroya and Kumar Ghosh, Sanjay and Toshniwal, Durga and Sekimoto, Yoshihide},
  journal = {Geoscience Data Journal},
  year    = {2022},
  doi     = {10.1002/gdj3.260}
}
@misc{arya2022rdd2022arxiv,
  title  = {RDD2022: A multi-national image dataset for automatic road damage detection},
  author = {Arya, Deeksha and Maeda, Hiroya and Kumar Ghosh, Sanjay and Toshniwal, Durga and Sekimoto, Yoshihide},
  year   = {2022}, eprint = {2209.08538}, archivePrefix = {arXiv}
}
```

### 2. UAV-PDD2023 — *supplement (aerial)*
2,439 UAV images, 11,154 instances, China, captured at 30 m.
- Source: Zenodo `8214118` (DOI 10.5281/zenodo.8214118) · Data in Brief 2023.
- License (per host): verify on Zenodo before reuse.

```bibtex
@article{uavpdd2023,
  title   = {UAV-PDD2023: A benchmark dataset for pavement distress detection based on UAV images},
  journal = {Data in Brief},
  year    = {2023},
  doi     = {10.1016/j.dib.2023.109692}
}
@dataset{uavpdd2023_zenodo,
  title     = {UAV-PDD2023},
  year      = {2023},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.8214118}
}
```

### 3. RoadDamageVision — *supplement (aerial/drone, China + Spain)*
Drone-captured road-surface images, manually annotated with bounding boxes for
six defect types (D00, D10, D20, D40, Repair, Block Crack); 7,647 instances,
D40 potholes most common (mostly the Spanish subset). COCO format.
- Source: **Mendeley Data `ypm4h4z25c` v3** (DOI 10.17632/ypm4h4z25c.3).
- **License: CC BY 4.0** (per Mendeley Data) — free to share/adapt with attribution.
- Authors: Luis Augusto Silva Zendron, Valderi Reis Quietinho Leithardt.

```bibtex
@dataset{roaddamagevision_mendeley,
  title     = {RoadDamageVision: Annotated Dataset of Road Damage Images},
  author    = {Silva Zendron, Luis Augusto and Leithardt, Valderi Reis Quietinho},
  year      = {2026},
  publisher = {Mendeley Data},
  version   = {3},
  doi       = {10.17632/ypm4h4z25c.3}
}
```

---

## Citing this compilation

Please cite **all three source datasets above**. You may additionally reference
this compilation by its Hugging Face repo URL. The maintainer of this
compilation claims no ownership over the underlying images or original labels.
# Unified Road Defect Dataset - Statistics


**Total images:** 30,186  |  **Total instances:** 75,765

## Instances per class

| id | class | train | val | total |
|---|---|---:|---:|---:|
| 0 | Longitudinal Crack (D00) | 27,114 | 4,807 | 31,921 |
| 1 | Transverse Crack (D10) | 16,358 | 2,992 | 19,350 |
| 2 | Alligator Crack (D20) | 9,941 | 1,769 | 11,710 |
| 3 | Pothole (D40) | 10,881 | 1,903 | 12,784 |

## Images per source

| source | train | val | total |
|---|---:|---:|---:|
| RDD-2022 | 20,215 | 3,552 | 23,767 |
| UAV-PDD2023 | 2,017 | 353 | 2,370 |
| RoadDamageVision | 3,445 | 604 | 4,049 |

_Manhole hard-negative backgrounds_: train 0, val 0

## Images per country / view

| country/view | images |
|---|---:|
| Japan | 7,900 |
| United_States | 4,805 |
| China_Spain_UAV | 4,049 |
| India | 3,223 |
| Norway | 2,914 |
| China_UAV | 2,370 |
| China_MotorBike | 1,934 |
| China_Drone | 1,919 |
| Czech | 1,072 |

## Image size (sampled 800)

- width:  min 512  median 640  max 4040
- height: min 512  median 640  max 2044