Performance Comparison of Yolo Versions for Small Object Detection on UAV Images

Abstract

With the rapid advancement in YOLO (You Only Look Once) versions, object detection has become a crucial task. In this article, the performance of three state-of-the-art YOLO versions: YOLOv8, YOLOv10, and YOLOv11 is analyzed. Using the VisDrone dataset, which contains 10 classes; these models were evaluated on metrics including precision, recall, F1 score, training time, and inference time. Based on the experiments, YOLOv8 achieved higher precision, recall, F1 score, and mAP compared to YOLOv10 and YOLOv11. On the other hand, YOLOv11 demonstrated reduced training time and inference time. Additionally, the performance of these detectors was analyzed on the Jetson Nano edge platform. YOLOv8 achieved the highest precision, recall, F1 score, and mAP on edge platforms as well. These experimental results provide valuable insights into selecting the most suitable YOLO version for specific object detection tasks and can guide further optimization efforts.