Bandung Pothole Image Dataset (BPID)

This dataset contains a collection of road infrastructure damage images, specifically potholes, directly acquired from various roads in Bandung City, West Java, Indonesia. It is specifically designed to represent dynamic, real-world environmental scenarios, making it highly suitable as an independent, out-of-distribution (OOD) test set to evaluate the robustness and generalization capabilities of Computer Vision models (e.g., Object Detection architectures like CNNs and Vision Transformers). The dataset comprises 50 annotated images capturing four distinct lighting and weather conditions: - Sunny: Maximum illumination on the asphalt surface, presenting visual challenges such as severe sun glare. - Cloudy: Diffuse and even lighting conditions with lower overall intensity. - Wet/After Rain: Wet asphalt surfaces and water puddles inside the potholes, which significantly alter the asphalt's visual texture and light reflection. - Night-time: Low-light conditions where object visibility relies heavily on artificial light sources (e.g., streetlights and vehicle headlights). Furthermore, this dataset captures specific infrastructure characteristics unique to the region, such as high-contrast tree shadows cast across the road surface. These visual anomalies act as hard-negative examples, which are particularly valuable for evaluating the False Positive rates of object detection algorithms.本数据集收录了实地采集自印度尼西亚西爪哇省万隆市多条道路的路面坑洼类道路基础设施损坏图像。该数据集专门针对动态真实世界场景构建，因此非常适合作为独立的分布外（out-of-distribution, OOD）测试集，用于评估计算机视觉模型（如卷积神经网络（Convolutional Neural Networks, CNNs）与视觉Transformer（Vision Transformers）这类目标检测架构）的鲁棒性与泛化能力。 本数据集共包含50张标注图像，涵盖四种典型光照与天气条件： - 晴天：沥青路面光照强度达峰值，存在强烈阳光眩光等视觉挑战。 - 阴天：光照均匀弥散，整体光照强度较低。 - 雨天/雨后：路面湿润，坑洼内存有积水，会显著改变沥青路面的视觉纹理与光线反射特性。 - 夜间：低光照环境下，物体可视性高度依赖人工光源（如路灯与车辆前大灯）。 此外，本数据集还收录了该区域特有的道路基础设施特征，例如横跨路面的高对比度树木阴影。这类视觉异常样本可作为难负样本，对评估目标检测算法的假阳性率具有极高应用价值。