wGrapeUNIPD-DL: An open dataset for white grape bunch detection
Marco Sozzi1, Silvia Cantalamessa2, Alessia Cogato3
1Department of Land Environment Agriculture and Forestry, University of Padova, Legnaro 35020, Italy.
Data in Brief
|July 25, 2022
Summary
This study introduces a new image dataset for computer vision in viticulture, featuring multiple grape bunches per image. This dataset enhances training efficiency for identifying grape clusters in vineyards.
Area of Science:
- Agricultural Science
- Computer Vision
- Viticulture
Background:
- Existing viticulture image datasets often lack multiple objects per image, limiting computer vision model training efficiency.
- Current grape variety catalogues primarily focus on individual plant structures rather than vineyard scenes.
Purpose of the Study:
- To develop a specialized image dataset for computer vision applications in viticulture.
- To enable more efficient training of object localization models for multiple grape cluster identification.
Main Methods:
- Acquired 373 images from vineyards in six Italian locations across various phenological stages.
- Captured images with a "later view" in vertical shoot position vineyards.
- Labeled images using the YOLO format, recording both actual and visible bunch counts.
Main Results:
- Created a dataset of 373 images suitable for multiple object identification.
- Provided images and corresponding labels in YOLO format.
- Collected data on the number of grape bunches in the field versus those visible in images.
Conclusions:
- The developed dataset addresses limitations of existing resources for computer vision in viticulture.
- The dataset facilitates improved training for models designed to detect multiple grape clusters.
- This resource supports advancements in automated viticulture technologies.
