Multi-class Road-Surface Condition Dataset for Automated Pothole and Road-Damage Classification

### Dataset Description This dataset is a multi-class road-surface image dataset developed for automated **road damage detection and classification in Bangladesh**. It contains **7,489 RGB crop images** derived from **3,154 original road images**. The source images were collected through two complementary approaches: **Google Earth imagery**, which provided broader coverage of road conditions, and **smartphone camera photographs** captured from real-world roads to represent practical and diverse surface conditions. Each visible road-damage region was annotated using **smart polygon-based instance segmentation**, providing detailed boundaries of individual damage regions. The polygon annotations were used to isolate each damage area from the original image, and the extracted regions were resized to **224 × 224 pixels** and stored as RGB images. This process produced one crop for each annotated damage instance. The dataset consists of **five road-damage categories and one background category**: * **Cracks:** Linear or irregular fractures appearing on the road surface. * **Damaged Asphalt:** Broken, worn, deteriorated, or spalled asphalt pavement. * **Open Manhole:** Exposed or uncovered manholes present on the road surface. * **Potholes:** Depressed or bowl-shaped holes resulting from pavement deterioration. * **Water-Filled Potholes:** Potholes containing standing water, which may partially obscure the damaged region. * **Background:** Clean and defect-free road-surface regions used as a rejection class. The dataset contains **1,689 Cracks, 1,830 Damaged Asphalt, 126 Open Manhole, 1,324 Potholes, 1,272 Water-Filled Potholes, and 1,248 Background samples**, totaling **7,489 images**. The dataset was designed to support **deep-learning-based road damage classification**, including individual CNN/transformer backbones and **multi-backbone ensemble models**. It can also be used for research on road-condition assessment, automated infrastructure monitoring, road damage severity analysis, and intelligent route recommendation systems.### 数据集说明 本数据集为针对孟加拉国自动化**路面损伤检测与分类**任务研发的多分类路面图像数据集。其包含7,489张RGB裁剪图像，源自3,154张原始路面图像。 原始图像通过两种互补方式采集：一是**谷歌地球影像（Google Earth imagery）**，可覆盖更广泛的路面状况；二是从真实道路拍摄的智能手机相机照片，用于体现实际且多样的路面状态。 所有可见路面损伤区域均采用**基于智能多边形的实例分割（smart polygon-based instance segmentation）**进行标注，可精准勾勒单个损伤区域的详细边界。通过多边形标注框从原始图像中分离出每个损伤区域，随后将提取的区域统一调整至224 × 224像素尺寸并存储为RGB图像，每一个标注的损伤实例对应一张裁剪图像。 本数据集包含5类路面损伤及1类背景类别： * **裂缝（Cracks）：** 路面上出现的线性或不规则断裂。 * **沥青路面破损（Damaged Asphalt）：** 破碎、磨损、劣化或剥落的沥青路面。 * **敞口检查井（Open Manhole）：** 路面上暴露或未覆盖的检查井。 * **坑槽（Potholes）：** 因路面劣化形成的凹陷或碗状孔洞。 * **积水坑槽（Water-Filled Potholes）：** 存有积水的坑槽，可能会部分遮挡损伤区域。 * **背景（Background）：** 干净无缺陷的路面区域，作为拒识类别。 本数据集共包含1,689张裂缝样本、1,830张沥青路面破损样本、126张敞口检查井样本、1,324张坑槽样本、1,272张积水坑槽样本以及1,248张背景样本，总计7,489张图像。 本数据集旨在支持基于深度学习的路面损伤分类研究，可适配单一卷积神经网络（CNN）/Transformer主干网络及多主干网络集成模型；同时也可用于路面状况评估、自动化基础设施监测、路面损伤严重程度分析以及智能路径推荐系统等相关研究。