Evolution of YOLO: A Comparative Analysis of YOLOv5, YOLOv8, and YOLOv10

Abstract

This paper presents a systematic comparative analysis of three versions of the YOLO (You Only Look Once) target detection algorithm - YOLOv5, YOLOv8 and YOLOv10. Through experiments on the VOC2012 dataset (converted to COCO format), this paper evaluates the versions in terms of multiple dimensions such as detection performance, inference speed and model complexity. The experimental results show that the detection accuracy and robustness significantly improve with version iteration and the mAP of v8 v10 is improved by 6.69% and 9.12% relative to v5, However, the number of model parameters increases by 68.98% and 48.66, and the FLOPS increases by 94.08% and 91.51%, respectively, which leads to an increased demand for computational resources and a slight decrease in inference speed compared to the old version, especially in practical application scenarios with limited resources.This paper not only demonstrates the continuous progress of the network structure and training strategy, but also explores the balance between performance and efficiency in real-time target detection, which provides references and insights for the future development of related technologies.