Model Assisted labeling

Label faster

Why model-assisted labeling?

As machine learning engineers, we understand the importance of high-quality labeled data for training accurate computer vision models. However, manually labeling large datasets to get ground-truth data is a time-consuming and labor-intensive task.

That’s where model-assisted labeling comes in. By leveraging your machine learning models right from the start, you can significantly speed up the data labeling process.

You can iteratively improve labeling accuracy while reducing the overall time and effort required

Set up efficient model-assisted labeling workflows easily on

1. Label the first point clouds or images

Images can be uploaded directly to through the web interface. Images and point clouds – single or in sequences – can also be uploaded via the URL when your data is in the cloud. This can be done via our API or Python SDK.

Once the images are uploaded, your in-house or external labeling team can get to work. You can take advantage of the many dedicated ML-assisted labeling tools to speed up the labeling.

Label 3D first. Then 2D.

There are i2 workflows for the data labeling. You can label the images first and then leverage those 2D annotations to speed up the 3D labeling. However, labeling in 3D space is usually more efficient, even if you only need 2D labels.

If your car drives by a stationary object like a traffic sign and captures it in 3 cameras for 100 frames, you have to annotate 300 2D bounding boxes.

In contrast, labeling the same traffic sign in 3D space only requires one cuboid annotation. Although labeling a 3D cuboid takes 3 times longer than a 2D bounding box, it’s still 100 times more efficient than labeling in images.

That’s why at, we prioritize 3D labeling and project to 2D only afterwards.

2. Train an initial model

Creating machine learning models for data labeling differs from models for vehicles or robots.

These data labeling models typically involve humans in the process and require interfaces that allow for effective interaction. Mistakes should be easy for humans to correct. We prioritize recall over high precision – it should be easier to delete a false positive detection than manually add a missing one.

Unlike other models, data labeling models do not need to be real-time, so we can optimize for accuracy even if it means sacrificing speed.

Finally, these models are allowed to cheat by considering future frames when making predictions on sequence data.

3. Predict labels on other images

Once your model is trained, it’s time to leverage its capabilities to generate label predictions for the remaining unlabeled images. Run your model on these images and upload the predicted labels to This step significantly reduces the manual labeling effort required.

multi sensor data labeling overlay and automation
data labeling accurate labeling merged view

4. Verify and correct predicted labels

After the 3D predictions are made, the next step is to manually verify and correct them. Humans are required for this task and need to be equipped with the right tools to interact with the data and machine-generated predictions efficiently. To make labeling point cloud data easier, we’ve developed several features.

The synced camera feature is the first one. It shows the camera image corresponding to where the mouse pointer is in the 3D space, making it easier to identify objects of interest.

Another handy feature is the batch mode, which allows quick adjustments to a particular object track’s cuboid throughout the sequence.

With ML-assisted cuboid propagation mode for labeling moving objects and a merged point cloud mode for labeling static objects that aren’t moving throughout the sequence.

These features make labeling more effortless. We’re also continuously working on new tools, such as the ability to automatically highlight potential labeling mistakes.

Workflow solutions is built by and for machine learning engineers. Fast, Accurate, and Efficient.

Autonomous robotics in a warehouse

Simplify multi-sensor labeling

Say hello to faster labeling, better quality data, and more time for your engineering team.
You’ve got access to a 14-day free trial to test it for yourself.