Investigation of YOLOv5 efficiency in iPhone supported systems

Abstract

Object detection gaining popularity and is more used on mobile devices for real-time video automated analysis. In this paper, the efficiency of the newly released YOLOv5 object detection model has been investigated. Experimental research has been performed to find out the efficiency of YOLOv5 using a mobile device with real-time object detection tasks. For this reason, four YOLOv5 model sizes have been used: small, medium, large, and extra-large. The experiments have been performed with a well-known COCO dataset. The original dataset consists of a huge number of images, so the dataset has been reduced to fit the mobile device requirements. The experimental investigation results have shown, that reducing the COCO dataset has no significant influence on the model accuracy, but the model performance is highly influenced by the hardware architecture and system where the model is used. Apple Network Engine usage might significantly increase the YOLOv5 model performance in comparison to CPU usage.