Automatic annotation

Magic Annotate finds every instance of one or more labels and draws a polygon around each one. Run it on the image you are working on, or across a whole image dataset in the background. To outline a single object with one click, use Magic Select instead (see Annotate images).

Before you start

• Magic Annotate works in image datasets, with edit access.
• It looks for objects by label name, so use category names such as person or car. See Labels.
• It uses credits, and starts only while your organization’s credit balance is positive. See Credit usage.

Confidence threshold

Every Magic Annotate run has a Confidence threshold, from 0.05 to 0.95 (default 0.5). Predictions below this confidence are discarded. Raise it for stricter results, or lower it to keep more candidates.

Existing annotations

When a selected label already has annotations on an image, you choose how to handle them. The choice applies only to those image and label pairs: every other pair is always processed.

Option	What happens
Skip	Keep the existing annotations and skip new predictions for that label.
Update	Replace that label’s annotations with fresh results.
Add	Keep the existing annotations and add the new results. Duplicates may occur.

Annotate one image

1. In the editor, select Magic Annotate in the toolbar (or press A).
2. Choose This image.
3. Select one or more labels (up to 64). To add a label, select Create label.
4. Set the Confidence threshold, and check the credit cost shown at the bottom of the panel.
5. Select Annotate. If some selected labels already have annotations on this image, choose Skip, Update, or Add, then select Annotate again.

The new annotations appear on the canvas when they are ready. If a label is not found, you see No instances of “label” found in this image.

All the selected labels run together as one prediction, so the cost does not grow with the number of labels. If Skip leaves no label to run, nothing is charged.

Annotate a whole dataset

1. On the dataset page, select Annotate > Magic Annotate. In the editor, you can also open Magic Annotate and choose Whole dataset.
2. Select one or more labels and set the Confidence threshold. Select Next.
3. Under Run on, choose Unannotated assets (images with no annotations at all) or All assets. Check the maximum credit cost, then select Next.
4. If some images already have annotations for a selected label, choose Skip, Update, or Add. The maximum cost updates with your choice. Select Annotate dataset.

If no image has annotations for the selected labels, the run starts as soon as you select Next in step 3.

Note

Before it starts, the wizard shows the maximum credit cost for your labels, scope, and conflict choice. Credits are charged per image as the run processes it. With Skip, images where every selected label already exists are not charged. If the balance runs out, the run stops; annotations already created stay.

Follow the run

The run continues in the background, and you can close the dialog. While it runs, the dataset header shows Annotating, and the dialog shows the progress: assets processed, annotations created, and how many predictions below the threshold were discarded.

Only one Magic Annotate run can be active per dataset. Wait for it to finish before you start another one.

Review the results

Annotations created by Magic Annotate are polygons, like any other annotation, and show their confidence score next to the label name. You can change their label, delete them, or reshape them with the area tools.

Magic Annotate also updates the review status:

• Images in To annotate that get new annotations move to Pending review.
• Approved or Rejected images that get new annotations move back to Pending review, so nothing skips your review.

To check the results, filter the dataset by Pending review and open the editor. See Review assets.

Next steps

• Review assets
• Annotate images
• Credit usage