Utah State's Vines-DB dataset trains AI to identify 7 ornamental vine species

A team from Utah State University—Saroj Burlakoti, Utsav Bhandari, Aaron Etienne, and Shital Poudyal—has released Vines-DB, a meticulously curated RGB image dataset designed for multi-species ornamental vine segmentation. The dataset comprises 1,218 original high-resolution photos of seven vine species (including Akebia quinata, Campsis radicans, and Wisteria floribunda) taken at the Utah Agricultural Experiment Station's Greenville Research Farm. All images were captured between July and October of 2023 and 2024 using an iPhone 16 Pro’s 48 MP camera, shot from a fixed 1m distance against black or white Styrofoam backdrops to reduce background noise. The 168 individual plants were grown on 1.2m x 2.4m trellises and photographed repeatedly across months to capture temporal changes in canopy development.

Each original image was manually annotated using Roboflow to create polygon-based instance segmentation masks for eight classes (seven species plus background). After preprocessing and data augmentation, the working dataset expanded to 2,307 images, split via stratified sampling into 2,019 training, 192 validation, and 96 test sets. The authors specifically chose to photograph between 10:00 AM and 12:00 PM under consistent daylight conditions to improve annotation reliability and model generalization. The dataset is hosted on OSF (DOI: 10.17605/OSF.IO/YJHCK) and is intended to support deep learning models for precision horticulture, urban ecology, and automated field phenotyping. Applications include canopy cover estimation, species identification, and benchmarking segmentation models under realistic outdoor conditions.