Viticulture Dataset: Grapevine Inflorescence Detection

Accurate detection of inflorescences can lead to an inflorescence count to provide an early yield estimation in viticulture. The task here is to detect inflorescences in RGB images using image processing, computer vision, and machine learning techniques. The purpose of this dataset is to provide a common platform for researchers to develop methods for inflorescence detection and compare their research.\nLineage: The RGB videos were captured for the harvest year 2018-2019 at two different vineyards in Woodside and McLaren Vale, South Australia. To this end, a GoPro Hero-5 camera was mounted on a ground vehicle to shoot the images/videos. The resolution of the images/videos is 4000×2976. The images/videos were captured in different weather conditions with variations in the background and brightness. The background variations include clear-sky, clouded-sky, and the sun, in the background. There was no customisation performed on the vines and the data collection was carried out in a contact-less manner. The inflorescences in the images/videos are completely visible or partially visible due to occlusions by leaves, stems, or other inflorescences. The inflorescences were labelled manually within the images using rectangular bounding-boxes. The images were labelled by the experts in the field of viticulture to avoid any kind of misjudgement in finding inflorescences in the images. An open-source software called ‘QuPath’ was used that provides tools to draw rectangles around inflorescences. A total of 558 images (80% for training and 20% for validation/testing) of size 4000×2976 were labelled.