PY-CrackDB: A Pavement Crack Dataset from Paraguayan Roads for Context-Aware Computer Vision Models

PY-CrackDB, provides a collection of high-resolution asphalt pavement images from Paraguayan roads, designed for training and validating computer vision models for automated infrastructure inspection. The data was collected from national routes in the vicinity of Coronel Oviedo, Paraguay, capturing the unique characteristics of local pavement types, environmental conditions, and fissure patterns. A primary contribution of this dataset is its focus on the detailed annotation of fine fissures (< 3 mm wide), a category critical for early-stage maintenance interventions as per Paraguayan road engineering standards. Methodology Images were acquired using two high-resolution smartphone cameras (Motorola moto g52 and Tecno Spark 10 pro) under standardized conditions. A ring light with a phone stand and a 55cm numeric ruler were used to ensure consistent lighting and scale at a capture height of 1.10-1.20m. All images have been processed and standardized to a resolution of 351×500 pixels. Manual annotation for segmentation was performed using Fiji software with a Wacom tablet, and a rigorous quality control process was implemented to ensure the accuracy of all labels. Dataset Structure The dataset contains a total of 569 unique scenes and is organized into two main directories, each corresponding to a specific computer vision task: Segmentation/ This directory contains data for semantic segmentation tasks and includes two sub-folders: Original_image/: Contains 369 original RGB images of asphalt pavement with cracks; Ground_truth/: Contains 369 corresponding binary mask images. In these masks, white pixels (value 255) represent the cracks, and black pixels (value 0) represent the background pavement. Classification/ This directory is structured for binary image classification tasks and includes two sub-folders: With_crack/: Contains 369 images that show pavement cracks. Without_crack/: Contains 200 images of pavement sections with no visible cracks. This resource is provided to support the scientific community in developing more robust and context-aware models for road asset management and safety.