AgriVision: A Benchmark Dataset for Advancing Real-World Robotic Vision in Densely Fruited Blueberry Crop

Area of Science:

• Agricultural Robotics
• Computer Vision
• Machine Learning

Background:

• Robotic vision for blueberry management faces challenges due to complex fruit structures, lighting variations, and cluttered backgrounds.
• A significant barrier is the lack of diverse, high-quality annotated data for training robust detection models in production environments.

Purpose of the Study:

• To introduce a large-scale dataset repository for dense blueberry analysis to advance learning paradigms in agricultural robotics.
• To address the data scarcity issue critical for developing effective blueberry detection models.

Main Methods:

• Developed a comprehensive dataset repository with three subsets: DB-1 (supervised learning), DB-2 (weakly/semi-supervised learning), and DB-3 (synthetic data).
• Proposed a data realization algorithm to generate realistic synthetic blueberry images, mimicking real-field complexity.
• Benchmarked a baseline model and proposed a customized framework for detecting densely clustered blueberries using the new dataset.

Main Results:

• The customized framework achieved 75.06% sensitivity (SEN), 56.85% intersection over union (IoU), and 72.49% Dice coefficient (DICE).
• Performance significantly outperformed the strongest baseline by 24.71% (SEN), 13.5% (IoU), and 8.6% (DICE).
• The synthetic data generation algorithm provides a scalable and cost-effective foundation for blueberry annotation and model generalization.

Conclusions:

• The introduced large-scale dataset repository is crucial for developing robust robotic vision systems for dense blueberry management.
• The proposed customized framework demonstrates superior performance in detecting densely clustered blueberries, paving the way for improved agricultural automation.
• The synthetic data generation method offers a valuable tool for overcoming data limitations in agricultural computer vision tasks.