Agreement-Based Labeling
  • 23 Oct 2023
  • 7 Minutes to read
  • Dark
    Light
  • PDF

Agreement-Based Labeling

  • Dark
    Light
  • PDF

Article Summary


LandingLens offers the ability to create and manage a Label Book. A Label Book is a living document to help guide users on how to properly label various categories of Classes. Now, you be asking:

  • How do we measure the consistency of labels across multiple users?
  • How do we ensure that only clean, accurate labels will be added to images?

The answer is Agreement-Based Labeling. Agreement-Based Labeling is the concept of allowing multiple users to label the same set of images to analyze the consistency of their labels. This method allows you to identify the areas where users may need more direction with labeling.

For example, let's say you have these Classes for Labelers to choose from:

  • Scratch
  • Rub Mark
  • Crack

Now take a look at the image below. Some Labelers may choose to label the identified area as a Scratch, while others may select Rub Mark or Crack. To help clarify this ambiguity, you can use Agreement-Based Labeling, then update your Label Book accordingly.

Ambiguous Mark to Be Labeled
Notes:
  • Classification projects support Labeling Tasks, but not Agreement-Based Labeling.
  • Agreement-Based Labeling is not applicable to Visual Prompting.

Agreement-Based Labeling Workflow

This is how Agreement-Based Labeling works:

  1. Send a Labeling Task to two or three users. This allows multiple users to label the same sets of images. 
  2. After the images have been labeled, LandingLens:
    • Calculates the similarities between the labels. This calculation is called the Agreement Score.
    • Ranks the images by Agreement Score, from lowest (least agreed and most confusion) to highest (most agreed and least confusion).
  3. You can then review the labels and accept or reject them. When you reject a label, you can provide feedback to the Labeler. To prevent issues in the future, add clarifying information to the Label Book.
    Agreement-Based LabelingAchieve Accurate Labeling with Agreement-Based Labeling

Labeling Tasks

Let's say you labeled some images, and you want to compare your findings with another user. To do this, you can send a Labeling Task to that user. 

LandingLens sends images in batches of 20 until all images have been reviewed by the required number of Labelers. However, if you send a Labeling Task with fewer than 20 images, LandingLens automatically divides those images into smaller batches. Let's explore how LandingLens sends images in an example.

Let's say you create a Labeling Task with these settings:

  • The Labeling Task has 30 images.
  • Two Labelers are required to label each image.

You then send the Labeling Task to these three Labelers: Anna, Cameron, and Eliza. The following steps will occur:

  1. LandingLens divides the 30 images into two batches:
    • Batch One: Includes the first 20 images 
    • Batch Two: Includes the remaining 10 images
  2. LandingLens sends Batch One to two of the Labelers. (Let's say, Anna and Cameron.) 
  3. Since two Labelers are required to label each image, the first person to complete their labels will be sent Batch Two. For example, if Cameron finishes labeling Batch One before Anna, then he will be sent Batch Two.
  4. LandingLens sends Batch Two to the third Labeler. (In this case, Eliza.)
    LandingLens Sends Images in Batches Until All Images Have Been Reviewed by the Required Number of Labelers 
Notes:
  • You can only send Labeling Tasks to Members who have been invited to your Project.
  • One set of images can only be added to one Labeling Task at a time; you cannot have the same set of images in multiple Labeling Tasks.

Send Labeling Tasks

  1. Open the Project.
  2. Select the images you labeled.
  3. Click Options in the action bar near the bottom of the screen and select Create Labeling Task. (You may need to scroll to see this option.)
    Create Labeling Task
  4. Enter a brief, descriptive name for the Labeling Task in the Task Name field.
  5. Select how many Labelers you want to label the images from the drop-down menu. For example, if you want two Labelers to label images, select 2.
  6. If you've already labeled some images, you can choose whether to display (carry over) those existing labels. Select No to remove the labels, or select Yes to display the labels. 
  7. Select the Labelers you want to send a Labeling Task to from the Assign Labelers drop-down menu. To add multiple users, click the + icon.
  8. Click Create Task. An email is sent to the Labelers, notifying them of the task.
    Create a Labeling Task 

Events That Occur After Labeling Tasks Are Sent

After a Labeling Task is sent, these events occur:

  • The New Labeling Task email is sent to the Labeler. This email includes a link to the requested images.
  • Images in a Labeling Task have the status In Task.
    The "In Task" Status Displays for Images in a Labeling Task
  • Images in a Labeling Task can only be labeled by the requested Labelers. (The requested Labelers can only label images using the link from the New Labeling Task email.)
    Images Can Only Be Labeled by the Requested Labeler  

Labeling Tasks for Segmentation Projects

Labeling Tasks in Segmentation Projects are assigned to users in batches of five images. For example, if you create a Labeling Task with 10 images and assign it to a user, that user will get two batches of five images each.

Labeling Tasks Are Grouped into Batches of Five Images

When the user opens a Labeling Task, they will see five images. When they submit that batch, the second batch of five images becomes available.

When You Finish a Batch, LandingLens Gives You the Next One

Limiting each batch to five images allows LandingLens to process each batch quickly.

Assign Additional Users to Labeling Tasks

  1. Open the Project.
  2. Click Tasks in the left side panel.
  3. On the Reviewing tab, click the vertical ellipses button and select Assign To.
    Assign To 
  4. Click the icon.
    Add a User
  5. Enter the name of the user you want to assign the Labeling Task to. (Remember, you can only invite users who have been invited to your Project.)
  6. Click Assign.
    Assign Another User to a Labeling Task

    Close Labeling Tasks

    You can close a Labeling Task before it has been completed. For example, if an assigned Labeler is unable to complete a Labeling Task, you can close it.

    To close a Labeling Task:

    1. Open the Project that has the Labeling Task you want to close.
    2. Click Tasks in the left side panel.
    3. On the Reviewing tab, click the vertical ellipses button and select Assign To.
      Close Task
    4. Click Yes in the pop-up window to confirm your action.

View and Submit Labeling Tasks

If a user has sent you a Labeling Task, you will receive the New Labeling Task email. 

To view an assigned Labeling Task:

  1. Click the link in the New Labeling Task email.
  2. The task opens. Label the images.
  3. After you've labeled all images, click Submit to share your labels.
    Submit a Labeling Task 

Review Submitted Labeling Tasks

After all the Labelers have submitted their Labeling Task, you will receive the Task [#] Has Labels Ready for Your Review email. This email includes a link to the labeled images. Click this link to compare the labels. You can also click the Review button on the Tasks page to compare the labels.

Review a Labeling Task 

The Review Task page with the labeled images will display. This page includes several settings to help you review the labels. The table below describes these settings. 

Review Task Settings 
NumberSettingsDescription
1Review Metrics
  • In Review: The number of labeled images that have yet to be reviewed.
  • Accepted: The number of images that have been manually accepted. 
  • Rejected: The number of images that have been auto-rejected or manually rejected. Images are auto-rejected based on the Agreement Threshold. (See number 3.)
2Hot Keys / Label Book / Submit
  • Hot Keys: LandingLens offers a variety of hot keys (or keyboard shortcuts) to help speed up the process of reviewing tasks. This button displays those hot keys.
  • Label Book: Opens the Label Book. This is useful so you can update the entries. 
  • Submit: After you've reviewed all the images, you can submit them to complete the review.
3Agreement ThresholdBy default, LandingLens calculates the similarities between all labels. This calculation is called the Agreement Score. The platform determines the Agreement Score based on the overlap of the Labelers' labels. For example, if the labels overlap by 50%, then the Agreement Score will be .50.
Note:
If there is only one Labeler assigned to a task, you will not see the Agreement Threshold.

You can adjust the required Agreement Score that images must have by changing the Agreement Threshold. For example, let's say the Agreement Threshold is set to .80. This means that the Labelers' labels must overlap by 80% to be approved.
Note:
A high Agreement Score does not necessarily mean that the labels are correct. It is still recommended to review these labels.
4Image AgreementThe Agreement Score for the opened image.
5SettingsThe Settings icon allows you to adjust the brightness and contrast of the image. This allows you to brighten images that were taken at night, for example.
6Labelers' LabelsBy default, the labels of all Labelers display on the image. You can select a Labeler to see their label only. This is useful if you want to accept the labels of a single Labeler.
7Agreement SortBy default, images are sorted by Agreement Score, from lowest (least agreed and most confusion) to highest (most agreed and least confusion). You can click the Agreement arrow to sort from the highest to lowest Agreement Score instead.
8Action buttons
  • Add Image to Label Book: There may be times when a Labeler is unsure how to label an image because the objects to identify overlap or are close together. You can add these types of images to the Label Book to help users in these rare cases.
  • Add Metadata: Tag images by adding metadata. Tagging images can help users locate specific images. Example tags: rare-case, ambiguous-case.
  • Review Notes: Leave feedback for the Labelers to help them understand why you rejected an image.
  • Accept / Reject: Accept or reject the labels of an image.
  • Edit: Update the labels.

Was this article helpful?