CloudFactory is excited to launch a new product that will give machine learning teams accurate image labels efficiently and quickly so they can get vision models into production faster.

In September, we shared how our acquisition of Hasty, a data-centric machine learning platform, would allow us to add industry-leading AI-assisted labeling to our established managed workforce.

That vision is now a reality with the launch of our new product: Accelerated Annotation.

This product blends the best of AI-powered automation and human expertise to deliver accurate labels for image datasets 5x faster. Accelerated Annotation lets you:

  • Reduce manual labeling while keeping human expertise in the equation.
  • Get the best result faster and with less effort.
  • Stop paying for labels you don’t need.

How Accelerated Annotation gets the job done

For over a decade, we’ve been known as one of the best-managed workforce providers in the world. We know how to get manual data labeling done and done well. Now, Accelerated Annotation is adding industry-leading AI-assistance to supercharge those capabilities.

Our stair-step workflow incorporates increasing levels of automation in an iterative process that accelerates labeling operations. We use automation where it works well, and where confidence is high while having access to an expertly trained and managed workforce to intervene where needed.

Standard Accelerated Annotation workflow

Standard Accelerated Annotation workflow

Here’s how Accelerated Annotation works:

Level 1: Out-of-the-box AI-Assistance

From the first image, our workforce uses out-of-the-box AI assistants to make labeling easier and faster. Level 1 assistants minimize the number of clicks required to complete an annotation by using neural networks pre-trained on generalized datasets.

These assistants are out-of-the-box and not yet customized to the use case, so human supervision in the form of clicks on regions of interest to include or exclude them in the annotation is required. This process is still powerful in shortening manual annotation, especially for instance and semantic segmentation, while the models learn and reach level 2.

Level 2: Custom Trained AI-Assistance

Once the custom models learn more about the specific use case, the labeling assistants begin suggesting annotations for data analysts to accept or modify. At this point, the workforce switches from annotation creation to a much faster, review and accept workflow stage. The labeling assistants are constantly retraining which minimizes required modifications or human supervision.

Annotation environment interface

Annotation environment interface

Level 3: Fully Automated

Level 3 automation is fully automated labeling. We auto-label batches of images with the custom-trained models powering the labeling assistants from level 2 automation. Importantly, we only migrate to level 3 when we see level 2 is working well. This way, we avoid the trap many teams fall into by forcing automation too early with unfavorable results.

Plus, AI Consensus Scoring and Active Learning

AI-assisted quality control and active learning are baked into each level.

AI consensus scoring technology leverages a machine learning method called Confident Learning, where a variation of AI models is used to find potential annotation errors in your dataset.

Our data analysts review potential errors, and those annotations that pass the AI consensus scoring review progress without further assessment. Think of this as a 100% quality assurance pass on every image and every annotation.

AI Consensus scoring interface where annotators can accept or reject AI labeling corrections

AI Consensus scoring interface where annotators can accept or reject AI labeling corrections

In addition to speeding the process up from an automated labeling perspective, our platform uses active learning to reduce the total number of manual labels required to reach full automated labeling. We leverage our state-of-the-art active learning approach to front-load a diversity of images when labeling a project. 

This ensures that the labeling assistant’s model shortens the learning curve - maximizing applicable automation. With this approach, we have reduced the number of images needing manual labeling by as much as 70%. This iterative, adaptive process ensures you are always focusing on the most relevant data.

Finally, a balance between quality and efficiency

Traditional approaches to data labeling have often fallen to extremes - manual labeling that does almost everything by hand or auto-labeling solutions that automate everything to a fault. Accelerated Annotation has finally brought together the best of both worlds, allowing companies to balance quality and speed through just the right amount of automation with an expert human touch.

We’re excited to help you innovate faster and more efficiently!

To learn more about our new product, visit our Accelerated Annotation page.

Data Labeling Computer Vision Data Annotation Tools Image Annotation AI & Machine Learning

Get the latest updates on CloudFactory by subscribing to our blog