An Analysis of Different YOLO Models for Real-Time Object Detection
Sawari Jamgaonkar, Jay Shyam Gowda, Siddharth Singh Chouhan, Rajneesh Kumar Patel, Ankur Pandey
Abstract
Object detection in real-life applications plays a significant role, with advancements in technology enabling enhanced image classification and processing, reducing manual effort and increasing accuracy and speed. The proposed work involves a comparative analysis and hyper-parameterization of YOLO (You Only Look Once) modules, specifically YOLOv7, YOLOv8, and YOLOv9 models. These models utilise Convolutional Neural Networks (CNNs) to process image layers by dividing images into boxes and then segmenting them into pixels for identification and classification based on various parameters. Initially applied to the COCO dataset, these models were further tested on customised datasets. Among the three, YOLOv9 demonstrates superior performance, achieving a mean Average Precision (mAP) of 52 at 0.5:0.95, compared to YOLOv7 and YOLOv8, which achieved mAPs of 35.2 and 39, respectively. Overall, YOLOv9 significantly reduces parameter usage and improves accuracy.