NEW
Say hello to our new Discord community
Join

Making farming bots better understand their environment through accurately segmented images

See how this company developing bots in the Farming industry used accurately labeled short range images to better understand the environment

 
Labeling short images for semantic segmentation for a farming automation bot

Challenge

The team needed to automatically identify and segment images for better scene understanding, so the bot could take appropriate steps dynamically.The camera produced high resolution images from a point of view close to the ground, which presented some difficulty in labeling the images correctly. In addition, the team needed to implement a tight QA workflow that was capable of producing highly accurate annotations while also being able to handle a large number of images.

Solution

The team had developed complex validation checks for annotated data as part of their data pipeline. To integrate this post-annotation step, their team used our SDK to automatically attach issues to the annotations that did not pass the validation checks. These were then checked by the annotation team on the platform and re-labeled if needed. In addition, the annotation team used our semi-automatic segmentation tool to quickly label objects accurately, massively improving the final labels.

Outcome

The ability to programmatically attach issues based on existing validation checks resulted in a massive reduction in the time spent by the team on QA. The semi-automatic segmentation tool reduced the annotation time by 20%, resulting in significant cost savings and reducing the TAT. This new workflow and the use of automated segmentation tool resulted in a jump in label quality from 91% to 96%, improving the accuracy of object detection models, resulting in more reliable arming bots.

A startup working on automation in the farming industry set out to develop Machine Learning models for object detection using a high-resolution camera. The goal was to automatically identify and segment images for better scene understanding, allowing the bot to make dynamic decisions based on its environment.
One of the main challenges stemmed from the camera’s perspective—positioned close to the ground - it captured high-resolution images from an unsual angle, that were difficult to label accurately. Ensuring high accuracy in segmentation while managing a large volume of images also required a robust quality assurance process to maintain high annotation standards.
To address these challenges, the team turned to Mindkosh’s data annotation platform, which is designed for highly accurate segmentation. Magic segment - our built-in semi-automatic segmentation tool proved especially valuable, allowing annotators to create smooth and accurate object boundaries with just a few clicks. This feature alone led to a 20% reduction in annotation time, resulting in cost savings and a faster turnaround time.
The customer's team had developed a complex data pipeline which included various validation checks, seamlessly integrated into the Annotation workflow. Using the SDK, the team automatically flagged annotations that failed validation, attaching issues programmatically, which could be viewed within the platform. Annotators could then review and correct these flagged annotations, ensuring that only high-quality data made it through.
Automating labeling images for segmentation with Magic segment
The customer's team had developed a complex data pipeline which included various validation checks, seamlessly integrated into the Annotation workflow. Using the SDK, the team automatically flagged annotations that failed validation, attaching issues programmatically, which could be viewed within the platform. Annotators could then review and correct these flagged annotations, ensuring that only high-quality data made it through.
A thorough quality assurance process further refined the dataset. Mindkosh’s quality control tools enabled spot-checking and correction of any inconsistencies, ensuring that the final labeled data met the rigorous standards required for autonomous applications. With a combination of automated assistance and rigorous validation, the annotation process became significantly more efficient.
By adopting this streamlined approach, the team not only improved annotation accuracy but also reduced the time needed to prepare training data. With a more efficient pipeline in place, the startup could accelerate its model development while maintaining the high-quality data necessary for real-world deployment.