MyUAV-UTHMPGR01: A CVAT-Annotated Drone-Captured Pothole Dataset and YOLO Benchmark for Aerial Road Inspection in Malaysia

Abstract

Road surface deterioration poses considerable safety and economic challenges in developing nations, prompting a growing demand for automated and scalable inspection solutions. Unmanned aerial vehicles (UAVs) provide a viable platform for aerial road monitoring; however, publicly accessible drone-captured pothole datasets tailored to Southeast Asian's road conditions remain notably scarce. To bridge this gap, this study introduces MyUAV-UTHMPGR01, a publicly available collection of 763 manually annotated aerial images acquired using a DJI Matrice 4E drone across road segments in Selangor, Malaysia. The raw images were annotated in the Computer Vision Annotation Tool (CVAT) in YOLO 1.1 format; the dataset is partitioned into training, validation, and test splits at the image level before any augmentation, ensuring a leakage-free evaluation protocol. Two lightweight detectors are benchmarked under matched conditions at $\mathbf{1024} \times \mathbf{1024}$ input resolution: You Only Look Once version 11 nano (YOLO11n) and YOLO version 8 nano (YOLOv8n). Under strictly controlled testing conditions, the two models demonstrated highly comparable performance. YOLO11n achieved a slightly higher mAP@50 of 0.461, edging out YOLOv8n at 0.455. To evaluate the models further, this study additionally trained and assessed both systems using the public HighRPD dataset, which features aerial imagery captured from an altitude of roughly 50 meters. In this scenario, YOLO11n maintained its advantage, scoring 0.361 compared to 0.325. However, the application of a zero-shot transfer test to the same dataset resulted in a severe drop in accuracy for both architectures. This sharp decline clearly illustrates the significant influence that camera altitude and image resolution have on object detection capabilities. These moderate values are consistent with the difficulty of small-target detection from UAV altitude and are reported as a transparent, reproducible baseline rather than an indication of model deficiency. The dataset, trained weights, and benchmark protocol are released to support reproducible research on aerial pothole detection.