A Comparative Analysis Of Yolov12 Model Sizes for Road Damage Detection on an Indonesian Dataset

Abstract

The sustainable maintenance of road infrastructure is a cornerstone of economic development and societal well-being. This thesis investigates the application of deep learning– based object detection for automated road damage assessment, focusing on a comparative evaluation of five YOLOv12 model variants ranging from nano to extra-large. The models were trained and tested on a large-scale dataset of Indonesian road imagery, which presents a challenging real-world environment with four distinct categories of road damage. The central aim of this research is to examine the trade-offs between detection accuracy and computational efficiency across different model capacities. Experimental results demonstrate a non-linear relationship between model size and performance, with the YOLOv12l variant achieving the highest accuracy. However, the relatively modest performance gap between the smallest and largest models indicates that lightweight architectures remain highly practical for deployment on resource-constrained edge devices. These findings underscore the importance of aligning model capacity with dataset characteristics and deployment requirements. Overall, this study contributes a benchmark analysis of YOLOv12 for road damage detection and highlights the broader implications of model efficiency in real-world AI applications.