To build a system that helps reduce the working hours of human apple pickers, an object detection model for detecting apples is essential. In this project, we create an apple detection benchmark with three different object detection models. Although
data augmentation generally improves results, our research shows that not all augmentations do. It also shows that using a model that has been trained on the MinneApple dataset and using it the RIWO dataset results in a decrease in performance. This research can be used as a stepping stone to continue the search of finding a usable and reliable object detection model for use in orchards. The YOLOv5X shows the best F1-score of 71%, and the second best, EfficientDet D3, shows a F1-score of 70% on the MinneApple dataset.
