Körber Supply Chain, a global leader in supply chain technology and automation, continually innovates to meet the complex demands of modern logistics. Specializing in optimizing warehouse operations, Körber develops advanced equipment for handling pallets loaded with diverse consumer goods. Until now, the equipment has mainly relied on a combination of non-vision sensors and manual settings from the operator to guide its operation, and occasionally, incorrect sensor readings, unexpected loads, or incorrect settings cause line stoppages, leading to downtime and lost revenue.
To address this, the Körber Supply Chain team decided to integrate cameras into their sensors and develop computer vision solutions to automate settings and manage multiple sensor errors that currently require manual inspection. The development of this computer vision system required building a large dataset of annotated images for model training. To annotate large volumes of data quickly, Körber turned to SuperAnnotate and managed to build datasets in a third of the time compared to their previous solutions.
“With SuperAnnotates AI-powered annotation tools, we can build our datasets in a third of the time compared to the alternatives we tried. Their easy integration with our cloud platform means we can focus on what matters, building cutting-edge logistics solutions faster, not worrying about software integration.”
-Emil Blixt Hansen, PhD and IoT Digital Developer at Körber
Körbers layer picker solutions
Körber's Layer Picker Solutions are designed to automate the handling of pallets in two primary use cases:
1. Pallet assembly: Creating pallets with mixed products by picking layers from different pallets onto a single one.
2. Pallet to tray: Depalletizing layer by layer and placing individual products onto conveyors for further processing.
Traditionally, these machines operated without vision systems, relying solely on sensors and manual configurations. This approach led to several challenges:
- Operator errors: With over 5,000 possible machine settings, operators often select incorrect configurations, resulting in inefficiencies.
- Product damage: Misconfigured settings or manufacturing defects sometimes cause products to fall or get damaged during handling.
- Sensor limitations: Sensors could be obstructed by hanging materials like cartons or plastic, causing unnecessary machine stops.
"We saw many incidents where goods would fall out of the machine because operators used the wrong settings," Emil explains. "We needed a solution that could not only detect these issues but also automate the selection of correct machine settings."
Improving reliability with computer vision
Körber introduced a computer vision system to the Layer Picker to address these challenges. They installed Intel RealSense depth cameras to capture top-down and side-view images of pallets before the Layer Picker picks the products. Using these images, they implemented object detection models to identify and classify products on the pallets.
This system then analyzes the pallet configurations and automatically selects the appropriate machine settings, reducing reliance on operator input. It minimizes mistakes and optimizes operations, ensuring products are handled correctly and efficiently.
"Our main idea is to calculate the correct picking program automatically," says Emil. "This will minimize mistakes and optimize our operations."
Choosing SuperAnnotate
While developing this computer vision system, Körber needed an efficient annotation tool for their machine learning models. Emil evaluated several options, including the Computer Vision Annotation Tool (CVAT). However, CVAT lacked user experience, integration capabilities, and automated annotation tools. In addition to spending more time configuring their data pipeline, each image took much longer to annotate as every product needed to be segmented by hand.
When the team tried SuperAnnotate, it quickly stood out for its advanced features, ease of use, and seamless integration with external platforms like AWS. The Magic Select tool, in particular, significantly sped up the annotation process, especially for images with hundreds of objects.
"SuperAnnotate's Magic Select tool significantly sped up our annotation process," Emil notes. "Also, we recently migrated to AWS, and the ability to integrate with AWS was crucial for our workflow."
Implementation with SuperAnnotate
Körber utilized two key features of SuperAnnotate to streamline their annotation process.
Efficient annotation with magic select and orchestrate
The Magic Select tool allowed for rapid instance segmentation by intelligently selecting objects with minimal manual input. This was especially beneficial when annotating images of pallets loaded with numerous items, such as cans, where manually drawing bounding boxes for each object would be time-consuming.
Additionally, the Orchestrate feature enabled Emil to create custom scripts that converted segmentation masks into bounding boxes, which were required for their object detection models. This automation further reduced annotation time and ensured consistency across the dataset.
"Annotating an image in CVAT could take up to 15 minutes," Emil recalls. "With SuperAnnotate, we reduced that time to about 5 minutes or less per image."
Körber recently transitioned to AWS, using AWS S3 buckets for data storage and AWS SageMaker to train their object detection models. The seamless integration with AWS SuperAnnotate allowed Emil to fetch annotations and integrate them directly into their training pipeline without additional servers or complex workflows.
Results and future plans
By integrating SuperAnnotate into their workflow, Körber achieved significant improvements:
- Improved annotation efficiency: Reduced annotation time per image by over 66%, freeing up valuable time for other development tasks.
- Model performance: Early models trained on approximately 500 images are showing promising results in automating the selection of machine settings.
- Operator independence: The system aims to reduce operator errors by automating settings selection, leading to increased operational efficiency.
"We're looking forward to being swamped with data as we start rolling it out to customers," says Emil. "The more data we have, the better our models will become, and SuperAnnotate will play a crucial role in labeling that data."
Scaling up
Körber is preparing for its first customer site test setup, which will provide a larger dataset for further training and validation. With the anticipated influx of data, Körber plans to enhance its models' robustness and accuracy. The team also intends to involve additional members in the annotation process, utilizing SuperAnnotate's collaborative features to manage the increased workload effectively.
Conclusion
Körber Supply Chain's integration of SuperAnnotate into their machine learning workflow has significantly enhanced their layer picking operations. The advanced annotation tools provided by SuperAnnotate have reduced annotation time, improved model accuracy, and increased operational efficiency by minimizing operator errors.
Interested in learning more about how SuperAnnotate can enhance your machine learning projects? Contact us for a demo!
