Wheat Head Detection using YOLO: A Comparative Study

Abstract

Classical object detection techniques such as RetinaNet, Fast R-CNN, and Single-Shot MultiBox Detector (SSD) are unable to locate objects in a single iteration. These approaches have solved the modelling and data scarcity issues in object detection. The YOLO algorithm has gained popularity because it performs better than the above-mentioned object recognition approaches. YOLO achieves cutting-edge findings and significantly outperforms prior real-time object identification algorithms by approaching object detection in a novel way. YOLO object detection algorithms advance from YOLOv1 [1] in 2016 attaining 63.4mAP on the Pascal VOC dataset to YOLOR in 2021 with 73.3 mAP on the far more difficult MS COCO dataset. This study assess the efficacy of the several YOLO algorithm iterations, from versions 3 all the way to version 7 and how well versions 3, 4, 5, 6 and 7 of the YOLO algorithms perform at identifying the wheat head from the provided object image.