Human labeling

Human labeling enables you to label images in your training dataset at scale. Well-labeled content results in better training data, which results in more accurate predictions from your model.

Google's human labeling service

You can use Google's human labeling service to label up to 5000 images per labeling task with a turnaround time that is typically one to five days. You can initiate a human labeling task directly from the AutoML Vision UI, and after this task is finished we will share labeled images for your approval before they are applied to your training dataset.

The service is currently free of charge. See the AutoML Vision pricing page for further details.

The requirements to initiate human labeling task are:

  • Your dataset has at least 100 unlabeled images
  • Your dataset has between 2 and 20 labels

To request human labeling:

  1. In the AutoML Vision UI, select the dataset containing the images you want to label.

    The Images tab shows the images in the dataset and the associated labels. If the dataset meets the requirements listed above (at least two labels, at least 100 unlabeled images, and so on), a message box appears near the top of the page offering the human labeling service.

    If the message box does not appear, you may need to add more images or labels or provide example images for each label.

  2. Click Request Labeling from the message box.

  3. Fill out the labeling request form that appears.

    You need to provide the human labelers with information about your use case, descriptions of your labels, example images for each label, and the number of images you want to have labeled.

  4. Click Submit Request.

    The system will begin processing your request.

Once human labeling is complete, the Images tab will show a Needs approval category in the left navigation between Labeled and Unlabeled. Select this category to review and approve the labels or take any other required action.

Was this page helpful? Let us know how we did:

Send feedback about...

Cloud AutoML Vision