Performance Validation of Yolo Variants for Object Detection
Kaiyue Liu, Haitong Tang, Shuang He, Yu Qin, Yulong Xiong, Nizhuan Wang
Abstract
Object detection is a core part of an intelligent surveillance system and a fundamental algorithm in the field of identity identification, which is of great practical importance. Since the YOLO series algorithms have good results in terms of accuracy and speed, YOLO and each subsequent version have been surpassing. Thus, in this paper, it carries out experiments on three versions of popular YOLO models such as yolov3, yolov4, and yolov5 (yolov5l, yolov5m, yolov5s, yolov5x). The performance of the three versions of YOLO model is analyzed and summarized by training and predicting the public VOC dataset. Results showed that the yolov4 model is higher than the yolov3 model in terms of mAP values, but slightly lower in terms of speed, while the yolov5 series model is better than the yolov3 and yolov4 models both in terms of mAP values and speed.