QuinceSet: Dataset of annotated Japanese quince images for object detection

Area of Science:

• Horticulture
• Agricultural Science
• Computer Vision

Background:

• Climate change necessitates adaptable fruit varieties.
• Visual assessment of fruit set is labor-intensive and requires expertise.
• Automated phenotyping can improve breeding efficiency.

Purpose of the Study:

• To introduce QuinceSet, an annotated dataset for Japanese quince (Chaenomeles japonica).
• To facilitate automated detection and phenotyping of quince fruits at different developmental stages.
• To support fruit breeding and yield prediction efforts.

Main Methods:

• Collected 1515 high-resolution RGB images of Japanese quince fruits at unripe and ripe stages.
• Manually annotated images using LabelImg software, creating 17,171 ground truth ROI annotations.
• Formatted annotations in YOLO format for machine learning applications.
• Acquired images under diverse conditions (weather, time, angle) and included partially obscured fruits.

Main Results:

• Developed QuinceSet, a comprehensive annotated dataset for Japanese quince.
• Dataset includes images from two key phenological stages: early fruit development and ripening.
• Annotations cover fruit detection and classification (unripe/ripe).

Conclusions:

• QuinceSet enables more efficient and reliable fruit breeding and yield estimation.
• The dataset can be utilized for automated identification and phenotyping of Japanese quince.
• This resource may also benefit breeding programs for other fruit crops.