Accurate fruit detection and counting are fundamental requirements in the development of reliable computer vision applications for yield estimation. This work was conceived to provide farmers with a farmer-friendly approach for automatic grape bunch detection. This study exploits the free demo version of the Roboflow 3.0 platform to train five state-of-the-art computer vision models with RGB images of white and red grape bunches, acquired with a smartphone in the field, and compares their performance. The results were evaluated both quantitatively, in terms of precision, recall, and AP@50 calculated on the validation set, and qualitatively on the test set. The models that achieved the best performances, also in the presence of overlapping clusters, were Roboflow 3.0 Object Detection and YOLOv11, reaching precisions of 86.6% and 88%, respectively, for the detection of white bunches, and of 85.7% and 89.9% for red bunches. This study highlights the possibility of developing highly accurate computer vision models for table grape bunch detection using the Roboflow platform, offering an accessible and user-friendly tool for non-expert users, including farmers.
Vicino et al. (Wed,) studied this question.