Label Assist
  • 05 Aug 2024
  • 3 Minutes to read
  • Dark
    Light
  • PDF

Label Assist

  • Dark
    Light
  • PDF

Article summary

This article applies to these versions of LandingLens:

LandingLensLandingLens on Snowflake

The Label Assist tool suggests new labels or improvements to labels after training a model in an Object Detection project. Use Label Assist to speed up the labeling process and help make your labels ("ground truth" annotations) more accurate, so that the next model will have better performance.

Label Assist can spot missing labels and suggest label improvements. For example, if there's a large space between an object and the original bounding box, Label Assist might suggest a tighter label.

Label Assist shows you its suggestions on the image details page as dashed bounding boxes. Click a suggestion to accept it, or click Accept All to accept all suggestions for an image. After reviewing and accepting suggestions, train the model again to include the newest labels in a model.

Label Assist

Label Assist Requirements

  • Label Assist is only available for Object Detection projects.
  • A model must be created for Label Assist to work. Label Assist offers suggestions based on the predictions of the selected model.

Use Label Assist

Follow the instruction below to turn on and use Label Assist. To see everything you can do with Label Assist, go to Actions Available in Label Assist.

  1. Open an Object Detection project that you've created a model for.
  2. Open the image you want to see suggestions for.
  3. If the Image Details panel isn't open, click the i button to open it.
  4. Click Label Assist in the Predictions panel.
    Turn on Label Assist
  5. Turning on Label Assist shows the suggestions and puts you in "label mode", which means that you can add and edit labels. 
  6. Each suggestion is marked with an animated dashed bounding box. Click a suggestion to accept it. This means that the suggestion turns into a label ("ground truth"). For example, in the screenshot below, clicking either of the purple dashed boxes will turn it into a "Hard Hat" label.
    Click a Suggestion (Animated Dashed Box) to Turn It into a Label
  7. You can also Click Accept All to turn all suggestions into ground truths.
    Accept All Suggestions
  8. You can edit any labels when Label Assist is on. This means you can add, resize, and delete labels.
  9. When you're done reviewing and accepting suggestions, click the toggle next to Label Assist.
    Turn Off Label Assist
  10. If you want to see Predictions again, click the Pan icon or press a to exit "label mode".
    Click "Pan" to See Predictions Again

Actions Available in Label Assist

You can perform the following actions when the Label Assist tool is on.

GoalActions and More Information
Accept a suggestionClick the suggestion (the animated dashed bounding box). 
Accept all suggestionsClick Accept All.
See different suggestionsMove the Confidence Threshold slider to show more or fewer suggestions. Higher thresholds typically yield fewer suggestions. Changing the threshold here doesn't change the model’s actual threshold.
Make suggestions more accurateSuggestions might not be completely accurate. For example, there might be too much space between the bounding box and the object of interest. In this situation, click a suggestion to accept it. Then resize and move the label.

You can resize and move all labels when Label Assist is on, including the labels that were added before the model trained.
Don't accept wrong suggestionsIf a suggestion is wrong, simply ignore it.
Turn off Label AssistWhen you're done reviewing and accepting suggestions, click the toggle next to Label Assist.

Run Label Assist on Images Uploaded After Training a Model

If you upload an image to a project after a model was trained, the model won't have any predictions for the image yet, which means that Label Assist won't have any suggestions. 

To get a suggestion on an image uploaded after a model was trained, turn on Label Assist and click Get Suggestion. This will cost 1 credit because the image makes a prediction on the image. (The credit cost is not applicable when using LandingLens on Snowflake.)

Get Suggestion

Was this article helpful?