Performance Analysis of YOLO26 in Pothole Detection on an Indonesian Road Dataset

Abstract

Road damage is one of the infrastructure problems that can compromise safety, comfort, and the smooth flow of traffic. The road inspection process, which is still carried out manually, requires a relatively large amount of time, labor, and cost, making a more efficient method necessary. Advances in computer vision and deep learning technologies enable the automatic detection of road damage through an object detection approach. This study aims to analyze the performance of the YOLO26l model as a baseline model in detecting four categories of road damage such as potholes, alligator cracking, lateral cracking, and longitudinal cracking using the Road Damage Indonesia Dataset. The dataset was divided into 70% training data, 15% validation data, and 15% testing data. The training process was conducted using the pre-trained weights from yolo26l.pt via the Ultralytics framework without any architectural modifications or the application of image enhancement methods. Performance evaluation was conducted using the Precision, Recall, mAP@0.50 (mAP@0.50), and mAP@0.50:0.95 (mAP@0.50:0.95) metrics. The results of the study show that the YOLO26l model achieved a Precision of 0.7028, a Recall of 0.6492, a mAP@0.50 of 0.6890, and a mAP@0.50:0.95 of 0.3391. Analysis using a confusion matrix, precision–recall curve, and visualization of the detection results showed that the model was able to identify all four categories of road damage well, although there were still some objects that went undetected under poor lighting conditions, due to small object sizes, or complex road surface textures. Based on these results, it can be concluded that YOLO26l performs well as a baseline model for road damage detection on the Indonesian road dataset