RoadDamageVision: Annotated Dataset of Road Damage Images

Research Hypothesis and Dataset Description

This dataset was developed under the hypothesis that deep learning and computer vision techniques can be effectively used to detect and classify various types of road surface defects from drone-based visual data. By providing annotated images captured in different countries and environments using aerial platforms, the RoadDamageVision dataset supports the development of scalable, low-cost, and automated road monitoring systems.

Data Overview and Collection Process

The dataset consists of annotated images of road surfaces showing visible damage. All images were captured using drones in both China and Spain, enabling aerial perspectives that provide a wide field of view and are suitable for large-scale infrastructure inspection. The imagery includes a variety of road environments, including urban, suburban, and rural settings.

Each image was manually annotated with bounding boxes and class labels for six types of road defects:

D00: Longitudinal cracks

D10: Transverse cracks

D20: Alligator cracks

D40: Potholes

Repair: Previously repaired surfaces

Block Crack: Block-type cracks

The annotations were standardized to facilitate training and evaluation of object detection models.

The dataset includes a total of 7,647 labeled instances of road damage. The most common defect is D40 (potholes), with 3,566 instances, primarily found in the Spanish dataset. In contrast, D10 and D20 appear only in the Chinese imagery. This difference in class distribution provides an opportunity to explore domain adaptation and class imbalance mitigation in model training.

研究假设与数据集说明
本数据集基于如下研究假设构建：深度学习与计算机视觉技术可有效依托无人机视觉数据，实现各类路面缺陷的检测与分类。道路损伤视觉数据集（RoadDamageVision）通过提供由空中平台采集、覆盖不同国家与场景环境的标注图像，为可扩展、低成本的自动化道路监测系统研发提供支撑。

数据概览与采集流程
本数据集包含带有可见损伤的路面标注图像。所有图像均由中国与西班牙境内的无人机采集，可获取具备宽视场的航拍视角，适配大规模基础设施巡检需求。该数据集覆盖城市、郊区与乡村等多种道路场景。

每张图像均已针对六类路面缺陷完成人工标注，标注内容包含边界框与类别标签：
D00：纵向裂缝；
D10：横向裂缝；
D20：鳄纹龟裂；
D40：坑槽；
Repair：已修补路面；
Block Crack：块状裂缝。

标注已实现标准化，以方便目标检测模型的训练与评估。

本数据集共包含7647个路面损伤标注实例。其中最常见的缺陷为D40（坑槽），共计3566个实例，且主要集中于西班牙采集的数据集内。与之相对，D10与D20仅出现在中国采集的图像中。这种类别分布差异为探索模型训练中的域自适应与类别不平衡缓解方法提供了研究契机。