End-to-End YOLO-Based Object Detection: From Model Training to Deployment

Abstract

This dissertation studies an end-to-end YOLOv8-based object detection workflow, from computer vision foundations and dataset preparation to model comparison and an illustra tive application prototype. Two experiments were conducted. The first experiment trained YOLOv8 variants for ship detection in maritime images, while the second used a custom 64-image egg dataset annotated manually with LabelImg. The models were compared using Precision, Recall, mAP@0.5, and mAP@0.5:0.95. In both experiments, YOLOv8 large was selected according to the stricter mAP@0.5:0.95 criterion, although the ship experiment also showed that YOLOv8 medium remained competitive on mAP@0.5 and Recall. The disser tation also presents GrabAndGo as an illustrative prototype using a pretrained YOLO-style detector to show how object detection can support cashierless shopping workflows. This application is included as an example of possible deployment, not as the main objective of the research or as a production-ready retail system. The work highlights the importance of dataset quality, annotation consistency, metric-based model selection, and careful system level framing when applying object detection models.