Kevric, ErminErminKevricWei, JiahuaJiahuaWeiBraun, Philipp MaximilianPhilipp MaximilianBraunRose, Hendrik WilhelmHendrik WilhelmRose2025-08-202025-08-202024-10IEEE International Workshop on Metrology for Agriculture and Forestry, MetroAgriFor 2024979-8-3503-5544-4979-8-3503-5545-1https://hdl.handle.net/11420/57082Developing systems for accurate apple detection plays a significant role in precision agriculture tasks and orchard management practices, hence having a direct impact on the economics of orchard fields. Thus, this research compares between different detection algorithms as it plays the baseline for subsequent tasks of counting, apple grasping and growth monitoring. In this work, four different state-of-the-art (SOTA) object detection models have been compared for apple detection: Faster R-CNN, YOLOv5, YOLOv8 and RT-DETR. The models were examined for their accuracy in apple detection and their invariance to the background apples. As Faster R-CNN and YOLOv5 are commonly used algorithms in this domain, they were chosen alongside YOLOv8. Furthermore, as the transformer-based detectors like RT-DETR outperform YOLO on large datasets, their comparison for apple detection is also warranted. The models are evaluated on the MinneApple dataset and a manually collected dataset. In addition, the effect of attention mechanisms in YOLOv8 architecture, hyperparameter optimization and data augmentation were explored. The evaluation showed that RT-DETR outperformed all models for mAP0.50:0.95 on the MinneApple and custom dataset for the trained image size. Furthermore, given the smaller background interference at a larger mAP, the RT-DETR model not only outperformed YOLO and Faster R-CNN models, but was also more background invariant.encomputer visionFaster R-CNNMinneApple datasetObject DetectionRT-DETRYOLOTechnology::600: TechnologyConference Paper10.1109/MetroAgriFor63043.2024.10948864Conference Paper