A Curated RGB Dataset for Real-Time Road Damage Detection using YOLO

This dataset consists of 1,075 high-resolution, annotated RGB images designed to support the development and benchmarking of real-time object detection models for road infrastructure maintenance. It is specifically curated to facilitate the detection of surface defects using the YOLO (You Only Look Once) architecture, contributing to automated pavement distress analysis and smart city monitoring systems. Dataset Composition The dataset is categorized into three primary classes of road distress, representing the most critical targets for structural assessment: ⦁ Crack: Longitudinal, transverse, or alligator cracking patterns in the asphalt surface. ⦁ Pothole: Significant structural cavities or depressions in the pavement. ⦁ Surface Erosion: General degradation, stripping, or localized wear of the top pavement layer. Technical Specifications ⦁ Total Images: 1,075 ⦁ Image Format: 3-channel RGB (Color) ⦁ Resolution: Harmonized to 640 X 640 pixels for optimal YOLO training efficiency. ⦁ Annotation Format: Provided in YOLO-compliant `.txt` format (normalized coordinates). Data Split and Preprocessing To ensure robust model evaluation and prevent overfitting, the dataset is partitioned into a training and validation framework: ⦁ Training Set (914 images): Used for supervised feature extraction and weight optimization. ⦁ Validation Set (161 images): An independent subset used for hyperparameter tuning and performance monitoring during training. ⦁ Preprocessing: Images have been resized and normalized to $640 \times 640$, maintaining aspect ratio through padding where necessary to ensure spatial consistency for the detector’s backbone. Applications This dataset is optimized for: ⦁ Real-time Object Detection: Development of lightweight YOLO models for deployment on edge devices or mobile inspection vehicles. ⦁ Pavement Management Systems (PMS): Automating the identification and quantification of road damage to prioritize repair schedules. ⦁ Edge Computing Benchmarking: Testing the inference speed and accuracy trade-offs of various YOLO versions (e.g., YOLOv8, YOLOv10, YOLOv12) on resource-constrained hardware.