Detecting grape bunches in vineyard images: a comparison of YOLO versions

Abstract

This work evaluates the performance of different YOLO (You Only Look Once) versions for grape bunch detection, focusing on the inference stage on heterogeneous hardware infrastructures. Devices ranging from a Raspberry Pi 3 and a mini-PC to mid- and high-end GPUs, as well as a conventional desktop CPU, were used. The analysis focuses on measuring the inference times, accuracy, and relative cost of each conf iguration, with the goal of identifying viable solutions under different budget constraints. The results highlight the limitations of low-resource systems and the remarkable performance of modern GPUs, facilitating the selection of the most suitable environment for precision agriculture applications. This study contributes to the development of accessible tools for automated crop estimation using the YOLO image object detection model.