Introducing Pre-labels Validation: A new way to label your data

We’re excited to introduce pre-labels validation, a powerful way to make your labeling more accurate, cheaper and faster by injecting your pre-annotations. This fundamentally changes the labeling process by letting our annotators focus on accepting, refining or fixing the pre-labels rather than annotating an image from scratch. 

Example use cases:

• Reviewing and correcting your model output 
• Changing annotation classes for a label schema change, e.g. breaking a dent class to big-dent and small-dent
• Improving an open source labeled dataset by treating the data as pre-labels
• Reviewing annotation output from your third party labeling teams

With each model iteration, the amount of labeled data required often increases exponentially — presenting a clear incentive for ML teams to increase labeling efficiency with pre-labels. 

Here’s how one of our customers used pre-labels during the feature’s beta phase:

How an automotive company used our pre-labels validation for 30% cost savings

A global renowned automotive company is training machine learning models to detect different car components for their next-generation smart factory. Using Bolt’s pre-labeling solution, the company was able to reduce their overall labeling costs by 30%.

We ingested pre-labels from their existing models for annotators to correct and/or adjust the classes and tightness of the annotations. There was no need to label the images from scratch – money saved for our client. The machine learning team then took the labelers’ output to improve the car components detection models further, which in turn generated better pre-labels for subsequent iterations. This helped to speed up the labeling process even more with each new pre-labels validation batch we did.

To start, simply select the Pre-labels validation project type while creating a new project.