Road Damage Dataset Evaluation Using YOLOv8 for Road Inspection System

The ASTM D6433-18 standard is widely used internationally in the Road Inspection System (RIS) to assess pavement distress, covering type, severity, and quantity. Automated detection of pavement distress types using vision-based methods follows these standards and requires a dataset with 19 types of pavement distress. The Road Damage Dataset (RDD) 2018 is a publicly available collection comprising 9,053 images captured on Japanese roads, each annotated with eight distinct types of road damage. However, only four of these classifications align with the categories specified in ASTM standards, namely alligator cracking, joint-reflection cracking, longitudinal and transverse cracking, and potholes. This article aims to assess the viability of utilizing the RDD dataset for Road Inspection System (RIS) purposes in accordance with the ASTM D6433-18 standard. The methodology involves the automated re-annotation of the dataset utilizing YOLOv8 models known as pseudo-labeling, followed by an evaluation to ascertain its compatibility with RIS requirements. The results suggest that the RDD-18 dataset is not suitable for conducting RIS in adherence to the ASTM D6433-18 standard. The evaluation results demonstrate less-than-optimal accuracy, which is attributed to an imbalanced distribution of instances among classes and a requirement for improved image quality. Finally, it is highlighted that the RDD 2018 dataset lacks images representing 15 additional types of pavement distress crucial for RIS applications based on ASTM standards.