Bandung Pothole Image Dataset (BPID)

Description

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.

Institutions

• Binus University

  Jakarta, Jakarta

Categories

Licence