Road Damage Dataset: Potholes, Cracks and Manholes

This dataset provides real-world road surface images collected in urban and rural areas around Rome and Sacrofano (Italy) using two low-cost devices: a GoPro HERO7 mounted on a moving vehicle and a Samsung Galaxy A14 smartphone for stationary captures. All images are 640x360.It aims to support realistic, device-independent road damage detection.

Each image is annotated in YOLO format with three damage classes:

Class 0 – Pothole

Class 1 – Crack

Class 2 – Manhole

Unlike most existing datasets that omit or misclassify manholes, this dataset explicitly includes them to improve model robustness and reduce false positives in real-world scenarios.

The dataset provides:

Polygon-based annotations (labels/): normalized quadrilaterals for each object (class_id, x_1, y_1, ..., x_4, y_4), suitable for oriented bounding boxes or segmentation tasks.

YOLO labels (labels-YOLO/): axis-aligned bounding boxes (class_id, x_center, y_center, width, height), ready for YOLO training.

COCO JSON annotations (annotations_coco.json): bounding boxes [x_min, y_min, width, height] in pixels, compatible with COCO-based frameworks.

Conversion scripts: YOLO-conversion-script.py and COCO-conversion-script.py to transform polygon annotations into YOLO or COCO formats.

The collection ensures diversity in lighting, perspective, and pavement texture, including urban, suburban, and rural roads. Images were captured at varying speeds (0–50 km/h), under different weather conditions, and with stabilized horizontal FOV. This variety enables reliable training and cross-device generalization for AI-based road inspection systems.