The authors of the QuinceSet: Dataset of Annotated Japanese Quince Images for Object Detection dataset address the pressing need for ecologically adaptable fruit varieties due to long-term temperature and weather fluctuations. The presented dataset comprises 1515 high-resolution RGB .jpg images, each measuring 3456×3456 pixels. Each image is associated with an annotated .txt file in YOLO format, marking the ground truth regions of interest (ROIs) for individual Japanese quince fruits. This annotation is meticulously performed with a total of 17,171 annotations provided by domain experts.

The authors highlight the significance of visually describing and evaluating fruit set characteristics at different growth stages for fruit breeding selection and yield predictions. This conventional approach, however, is labor-intensive and time-consuming, requiring both human expertise and effort. To overcome these challenges, the authors introduce the QuinceSet dataset, which comprises images of Japanese quince (Chaenomeles japonica) fruits annotated for detection and phenotyping across two phenological developmental stages: post-flowering when the fruit reaches 30-50% of its final size and ripening stage just prior to yielding.

The dataset is collected from the Institute of Horticulture in Dobele, Latvia, capturing images of both fully visible quinces and those partially obscured by leaves. To ensure data reliability and quality, images are homogenized under varying weather conditions, different times of day, and various angles of capture. The images were collected in field conditions from different Japanese quince genotypes, characterized by distinct shrub forms and fruit shapes. Captured with a Samsung Galaxy A8 cell phone, the images were taken from an orchard plot at the Institute of Horticulture.

Prior to image capture, experts from the Institute of Horticulture evaluated breeding conditions and optimal imaging timing for Japanese quince. Images were acquired under diverse weather conditions, varying distances, and angles. The dataset is divided into two phenological stages: unripe and ripe quinces. The former is captured post-flowering, once the second fruit fall is complete and the fruit reaches 30-50% of its final size. The latter corresponds to the ripening stage just before fruit yield, with data collected at varying times to account for genotype differences.

Image annotation was meticulously executed using LabelImg software, associating each image with its corresponding class and class-specific bounding box coordinates in YOLO format. Annotations were performed to encompass the entire quince fruit, occasionally leading to overlapping annotations to ensure comprehensive coverage. The authors’ dataset contributes to more efficient breeding processes, robust yield estimation, and enhanced phenotyping of quinces, while potentially offering insights for the breeding of other crops as well.