IRDD: Iraqi Road Damage Dataset for Deep Learning-Based Road Damage Detection

The Iraqi Road Damage Dataset (IRDD) is a novel image dataset developed tosupport deep learning-based road damage detection under Iraqi pavement andenvironmental conditions. It was created to address a significant domaingap identified when applying models trained on international benchmarks(e.g., RDD2022) to Iraqi roads, where preliminary experiments showed asubstantial drop in detection confidence and a high non-detection rate. The dataset consists of 8,479 images collected via smartphone dashcamacross road networks in Wasit Governorate, Iraq. All images were manuallyannotated using the Oriented Bounding Box (OBB) format across four damageclasses, consistent with the RDD2022 taxonomy: D00 (Longitudinal Crack),D10 (Transverse Crack), D20 (Alligator Crack), and D40 (Pothole). The dataset underwent a rigorous automated quality control process,including image integrity validation, label format verification, classID validation, and coordinate range checks, ensuring zero invalidannotations in the final release. Data is provided pre-split intotraining (5,935 images), validation (1,271 images), and test (1,272images) sets using iterative multi-label stratification to preserveclass balance across all splits. This dataset is intended to support research in road infrastructuremonitoring, pavement distress classification, domain adaptation forcomputer vision, and smart city applications, particularly in regionswith underrepresented road conditions in existing public datasets. Folder structure:- train/images, train/labels- val/images, val/labels- test/images, test/labels- data.yaml (YOLO-format configuration file) Annotation format: YOLO OBB (class_id + 8 normalized coordinates per line)