Fruitlet image dataset for apple phenotyping during early development

This dataset originates from an experimental study conducted between April 24 and May 29, 2024, aimed at providing agronomists with an automated vision tool to accelerate data collection of apple fruitlets during early development. Excluding corymbs that experienced total fruit abscission, the data acquisition process resulted in 234 video files in .bag format and a total of 1,054 fruitlet measurements. The dataset includes three primary data sources:

bag_videos.zip: a collection of videos recorded using the Intel® RealSense™ Depth Camera D435i. Each video captures target fruitlets from multiple orientations, with an average duration of 10 seconds, at a distance of approximately 30 cm, and a resolution of 640 × 480 pixels;

ground_truth_caliper_measurements.csv: the corresponding ground-truth measurements of selected corymbs, collected across 7 monitoring sessions. Measurements were categorized by date and bud type to analyze growth differences over time. Metadata such as the orientation of the vegetative wall and the presence of the king fruit was also recorded;

FruitletDetectionDataset.zip: a dataset for model training, validation, and testing, comprising 481 images and corresponding oriented bounding box annotations. The images were obtained through stratified random sampling after RGB frame extraction, ensuring balanced representation across videos. The dataset is split into training (60%), validation (20%), and test (20%) subsets.

The code to process and reproduce these data is available in our GitHub repository.