Comparative Analysis of YOLOv8 and Other Object Detection Models Using the COCO Dataset

Abstract

Object detection is now an integral element of contemporary computer vision, which helps machines to decipher image data for uses from autonomous driving to intelligent surveillance systems. In this research, we perform an extensive comparative assessment of YOLOv8, the newest member of the YOLO family, compared to other best-in-class object detection models, i.e., YOLOv5, Faster R-CNN, RetinaNet, and DETR, across the COCO dataset. The performance of models is assessed using performance measures like mean Average Precision (mAP), inference speed, and the ability to qualitatively detect. It is evident from our findings that YOLOv8 strikes an excellent trade-off between speed and precision and thus outperforms conventional models, especially for applications requiring real-time output. The conclusion provides useful observations to the extent that they shed light on how models may be selected for realistic applications where various needs are catered to. YOLOv8 is the most effective model that scores the highest mAP value of 0.48 at the last epoch and exhibits its capacity to quickly reach convergence and surpass other models. Even though DETR holds the highest accuracy at initialization, which is outmatched by YOLOv8, it also lacks better convergence and larger inference times compared to YOLOv8, which is its deprecating feature for use for applications that rely on low-latency and high-accuracy applications. The comparison pinpoints YOLOv8’s advantageous trade-off between accuracy and inference speed and thus makes it better-suited for deployment on low-latency systems, especially applications demanding low-latency and high-accuracy.