Automate annotation | Viam Documentation

Automate annotation

Manually labeling hundreds of images is slow. If you already have a trained ML model – even a rough first version – you can use it to generate label predictions automatically. You review and correct the predictions instead of labeling from scratch.

You can also use the SDKs to annotate images programmatically, or combine capture, annotation, and dataset management into a single script for continuous dataset improvement.

Auto-predict labels

Use an existing ML model to generate predictions for images in a dataset, then review each prediction. This works with both classification models (which generate tags) and object detection models (which generate bounding boxes).

  1. Navigate to your dataset’s page.
  2. Click Get auto-predictions.
  3. Select a model to generate predictions with.
  4. Set the confidence threshold above which to create a label prediction.
  5. Click Get predictions.
  6. Once predictions have finished generating, click Review predictions.
  7. For each image, Accept (A) or Reject (R) each prediction.

Programmatic tagging with SDKs

Use an ML model to generate tags for images, then pass the tags and image IDs to the data client API.

Programmatic bounding boxes with SDKs

Use an ML model to generate bounding boxes for images, then pass each bounding box and image ID to the data client API.

Capture, annotate, and add to dataset in one script

The following example captures an image, uses an ML model to generate annotations, and adds the image to a dataset – all in a single script. Use this pattern to expand and improve your datasets continuously over time. Check annotation accuracy in the DATA tab, then retrain your ML model on the improved dataset.

What’s next