Model-Assisted Labeling

Reduce annotation time and get a model into production faster by using model weights from a trained model to label your data
In February 2021, we released model-assisted labeling. Model-assisted labeling allows you, with one click, to have a model take a first pass at annotating your images. While we strongly encourage that a human reviews the annotations to make sure the labels adhere to best practices, this feature significantly speeds up your team's annotation efforts!
Label Assist with the Microsoft COCO Model
Our Public (Community), Sandbox (Business Proof of Concept), Growth, and Enterprise accounts support using the predictions from your Roboflow Train models, our trained version of the Microsoft COCO 2017 Dataset for Object Detection, or a trained model from Roboflow Universe as a starting point for annotating additional images.

Label Assist with Microsoft COCO Model Weights

Label Assist with Trained Universe Models

Check that you are operating in the same workspace on Universe where you will want to use the model. You can do this by clicking on your Workspace Profile Name in the top-right corner of the Universe header and clicking "Switch Your Workplace". Then select the workspace in which you will use the model, and click "Select" to confirm the change.
Changing Your Universe Workspace Profile
Confirming Your Universe Workspace Profile Selection
Go to Roboflow Universe and search for a model that you want to use with Label Assist.
Then, click the “star” icon next to the project name for a dataset with a Model tag on it. Here’s how the star appears on a project card in a search page:
Starred Dataset (Model Tags are colored in Sea foam Green)

Activating Label Assist

Navigate to the annotation tool on a project that you are working on.
Click on the Annotation Batch you'd like to label.
Select an image in the Annotation Batch, and click the Label Assist icon (the Magic Wand) in the Labeling UI to activate Label Assist.
Click “Public Models”. In the “Project” dropdown, you will see the projects you have starred.
Click on the model you want to use to assist you while labeling. Then, click “Continue”.
After a few moments, you will be asked to remap labels if needed to ensure that the pre-trained model uses the right labels to match the class names in your project.
If you remap the label(s), click the green check mark next to the label name once you are satisfied with the new label name(s).
Select "Let's Annotate!" and use the Left/Right arrow symbols in the top section of the Labeling UI to navigate to the next image, or use the Left/Right arrow keys on your keyboard.
Adjust the Confidence and Overlap sliders to fine-tune your automated labeling process.
Select "Disable Assist" when you're done labeling, or click the back-arrow next to the Project Name (e.g the arrow next to "Pills" in the example below) when you're finished.
Label Assist with a trained Roboflow Universe model

Next Steps

Generate a dataset version, and use Roboflow Train to train your model (without having to spin up and manage your own training infrastructure).
Alternatively, you can export your dataset and use it with the Model Library (or your existing training infrastructure) to train a model.
  • NOTE: If you use Roboflow Train, then you'll have immediate access to deployment through the Deploy Tab, inference via API, and to edge devices.
  • If you choose the custom training route, we have a GitHub repository with written and video tutorials, as well as pre-written training notebooks, to help you get up and running.