Comparison of Segmentation Models for Pothole Detection Using UAV Imagery

Abstract: Potholes are a significant cause of vehicle damage and road safety accidents. Existing inspection practices rely on manual surveys using inspection vehicles or vehicle-mounted cameras, which are time-consuming, non-scalable, and often provide incomplete visibility. This paper presents a comparative study of instance-segmentation-based deep learning models for pothole detection using UAV imagery, enabling wider and more consistent coverage. A custom dataset was developed and augmented using geometric transformations. YOLOv8 and transformer-based RF-DETR models were evaluated using precision, recall, and mAP@50. The best model achieved mAP@50 of 66.3%, highlighting challenges in detecting shallow potholes under limited training data.