Benchmarking YOLOv8 Variants for Automated Infrastructure Assessment

Abstract

Structural cracks in buildings and bridges pose a serious safety concern in Nigeria, with manual inspection methods being slow, subjective and impractical at scale. Undetected structural cracks can lead to building collapses, which have become a recurring and deadly problem in Nigeria's construction sector. This study benchmarks three YOLOv8 variants – nano, small and medium under identical training and evaluation conditions using a 500-image Building Crack Detection dataset sourced from Roboflow Universe, with models evaluated on mAP@50, mAP@50-95, Precision, Recall and Inference Speed. All three models were trained for 50 epochs under the same conditions to ensure a fair comparison across variants. All three models achieved a mAP@50 of at least 99.4% and 100% recall, with YOLOv8s and YOLOv8m both achieving the highest mAP@50-95 at 99.50% and YOLOv8n achieving the fastest inference speed at 3.1 ms per image. These results show that all three models were highly accurate and rarely missed a crack, with the main difference between them being how fast each one could process an image. YOLOv8n was identified as the most optimal model for Nigerian infrastructure deployment, demonstrating the best balance of accuracy and speed for resource-constrained settings. Although the larger models were slightly more precise, the accuracy gap was small enough that speed became the deciding factor for practical deployment. This study demonstrates the viability of automated crack detection using YOLOv8 and provides a foundation for affordable and scalable structural monitoring in Nigeria.