How Oishii Uses Labellerr to Train Strawberry-Harvesting Robots
Oishii, a vertical farming innovator, needed to train robots to detect strawberry ripeness and stem cutting points. But challenges like overlapping berries, shadows, and inconsistent lighting made annotation complex. High-accuracy segmentation of berries, calyx, and stems was critical and internal teams lacked the bandwidth to handle it at scale.
Partnering with Labellerr led to 70% faster annotation, pixel-perfect segmentation, and better AI model performance. With bounding boxes, polygon masks, and keypoints integrated into Oishii’s workflow, the robots could reliably identify ripe berries and cut accurately, reducing damage, waste, and manual labor in high-density vertical farms.
Introduction
AI is changing industries globally, from healthcare to logistics. In agriculture, AI helps with labor shortages, crop quality, and harvesting issues.
In vertical farming, AI and robotics are even more crucial because they help farmers make the most of limited space and controlled environments.
Automated harvesting robots need to detect strawberry ripeness and find precise cutting points on the stem.
Accurate image segmentation is vital for training these robots, forming the foundation for their performance.
Our Customer
Oishii is a vertical farming company that uses robots to make farming smarter and more efficient.
They use AI and robotics to solve farming challenges, aiming to maximize production, reduce manual labor, and ensure precise harvesting.
For this project, Oishii aimed to develop an AI model for autonomous robots to identify and harvest strawberries with high accuracy.
This precision helps robots harvest strawberries efficiently without damaging the fruit or plant, making strawberry farming more sustainable and scalable.
The Complex Data Labeling Challenge
Oishii's project required highly detailed and accurate image labeling for their AI.
- Data Complexity: Labeling strawberry images in vertical farming was complex. Annotators had to carefully label berries at different ripeness levels (flower, transition, unripe, ripening, ripe).
Strawberries often overlapped or were hidden by leaves, making them hard to identify. Artificial lighting also created shadows and reflections that affected how berries, stems, and calyx appeared in images. - Volume of Data: The project involved thousands of images, each containing multiple strawberries needing detailed labeling. This meant tens of thousands of data points to label. Managing such a high volume while keeping consistency and quality was a big challenge.
- Accuracy Requirements: High accuracy was critical because the labeled data guided the robot. Any errors could lead to cutting unripe berries or damaging plants, reducing efficiency and increasing waste. Precise and consistent labels were essential for the AI model to work effectively.
How Labellerr Stepped In
Labellerr customized its platform to meet Oishii’s specific needs for strawberry labeling.
1. Tailored Solution and Annotation Types
Labellerr introduced a grouping tool to link all labels for one strawberry (berry, calyx, and stem) as a single unit, helping the AI understand their relationship. Labellerr used various annotation types:
- Bounding Boxes: These marked the overall shape and location of strawberries to detect ripeness.
- Polygons (Masks): These outlined the exact edges of berries, calyx, and stems, allowing the AI to differentiate them accurately. Annotators drew masks only on visible areas, splitting them if an object obstructed the view.
- Key Points: These were placed on specific parts of the stems (2mm above the calyx) to show precise cutting points for the robot. For each berry, four points marked width and height, and one point marked the stem.
2. Advanced Features
Labellerr's platform included advanced features to make labeling faster and more accurate:
- AI-Assisted Annotation: The platform used AI to pre-label data, suggesting bounding boxes, polygons, and key points. This reduced manual work and speed up the process.
- Quality Assurance Mechanisms: Features like confidence scoring helped identify areas needing review, and anomaly detection flagged inconsistencies, ensuring high accuracy and uniformity.
3. Workflow Integration
Labellerr's platform seamlessly connected with Oishii's existing data systems. It supported easy uploading of images, efficient labeling, and exporting of data in formats compatible with Oishii's machine learning workflows. This integration allowed Oishii to process data smoothly and use it in their AI development.
Results and Impact
The collaboration between Oishii and Labellerr delivered significant improvements:
1. Improved Annotation Efficiency
Labellerr automated repetitive tasks, speeding up the labeling process significantly. Oishii completed annotations in 70% less time than manual methods. Tasks that used to take weeks now took only days, saving time and money.
2. Enhanced Model Performance
Labellerr's precise labels greatly improved the AI model's performance. The robot could accurately detect strawberry ripeness and identify exact stem cutting points, making harvesting more reliable.
This meant the robot avoided cutting unripe berries and damaging plants, improving harvest quality and reducing waste.
3. Scalability
Labellerr's tools allowed Oishii to handle large datasets easily, preparing them for future needs. The platform managed thousands of images efficiently, ready for expanding to other crops or advanced applications.
Conclusion
Labellerr’s automated labeling and quality control capabilities greatly helped Oishii develop a reliable AI model for robotic strawberry harvesting in vertical farming. By making labeling easier, ensuring precise segmentation, and saving time, Labellerr helped turn Oishii’s vision into a successful AI-powered solution that improves efficiency and productivity in vertical farming.
.png)
