A Novel Dataset for Baby Broccoli Identification by using YOLOv8 Model

As global population increases, the need for agricultural automation becomes crucial for a stable food supply. An autonomous mechanism for picking baby broccoli could economically benefit farmers and society. To date, efforts to develop an automated baby broccoli harvesting solution have been absent. A key step in automation, particularly for the automated recognition of baby broccoli heads through computer vision, involves the creation of a rich dataset that sufficiently captures the characteristics of baby broccoli heads, as such a dataset has not yet been collected or published. This paper marks the first step towards automating baby broccoli harvesting by creating a novel dataset and testing its accuracy for precise identification. We have gathered data using custom software and a RealSense D435 Depth camera, known for its depth and stereo vision capabilities. The initial results of the model that was trained from our dataset had mAP values ranging from 0.869 to 0.942. This initial training shows that the dataset fits the purpose of detecting baby broccoli heads. Further results shows that when the model is 97.7% confident or more, its predictions are 100% precise.