Skip to main content

Annotator Guide

Roles and Prerequisites​

Target Roles​

RoleDescription
AnnotatorExecutes annotation, submits for check, and handles unqualified results
Administrator, Project ManagerPerform the same status operations and additionally can edit and delete tasks
Auditor, CollectorDo not perform annotation

Prerequisites​

ItemRequirement
AccountCreated by an administrator with the Annotator role assigned; already added to the project that owns the task
Task assignmentAssigned as the annotator of the target task
PermissionsView Annotation Tasks and Perform Annotation
Task statusThe status is Pending Start, In Progress, Pending Check, or Unqualified
DictionaryThe annotation dictionaries for skills, objects, targets, and other fields are maintained by an administrator

Sign-in page

Task Statuses and Available Actions​

The task state machine has six statuses. The table below gives the meaning of each status and the status operations available to an annotator.

StatusMeaningAnnotator can setNext status available to the annotator
Pending StartThe task is created but not yet startedYesIn Progress
In ProgressAnnotation is under wayYesPending Check
Pending CheckAnnotation is finished and awaiting checkYesIn Progress
UnqualifiedThe check failed and rework is requiredNoPending Check
QualifiedThe check passedNoNone
Submit DataThe data is submitted and the workflow is finishedNoNone

An auditor sets Unqualified; an annotator cannot move a task to that status. When a task is Unqualified, the annotator can set it back to Pending Check. When a task is Qualified, the annotator cannot change its status.

Operation Workflow​

Opening the Annotation Interface​

  1. In the Data group on the left, select Annotation to open the annotation task list.
  2. Open the task assigned to the annotator to enter its details page.
  3. On the Annotation Data tab, select or click the target dataset.
  4. Select Start labeling; if the dataset already has annotations, the button reads Continue Annotation.
  5. The annotation interface opens in a new browser tab.

The task details page has four tabs.

TabPurpose
Annotation DataList of datasets linked to the task; the entry to start annotating
Batch AnnotationQuick Annotation, Batch Copy Annotations, and Batch Replace Annotations
Problem annotationList of annotations marked as invalid
Annotation StatisticsAnnotation descriptions and statistics

Annotation interface

Annotating and Saving​

  1. In the annotation form, select the skill, object, target, and adverbial. If nothing matches, add a custom dictionary item.
  2. Press Q to set the action start time and R to set the action end time.
  3. Press Enter to add the annotation.
  4. Press Ctrl+S to save the annotation.
  5. Repeat these steps for the remaining actions in the dataset.

The annotation interface uses the following shortcuts.

ShortcutAction
QSet start time
RSet stop time
EnterAdd annotation
Ctrl+SSave annotation
SpacePlay or pause
← / →Seek backward or forward

Before submission, the form validates the following items.

MessageMeaning
The start time is not setThe start time was not set with Q
The end time is not setThe stop time was not set with R
The start time cannot be greater than or equal to the end timeThe time range is reversed
The skill is not setNo skill was selected
The object is not setNo object was selected

Annotation interface

Batch Annotation​

The Batch Annotation tab on the task details page offers three modes.

ModeBehavior
Quick AnnotationAdds the same annotation description to every dataset in the task
Batch Copy AnnotationsCopies a sample annotation to other datasets in proportion to the timeline
Batch Replace AnnotationsReplaces descriptions, objects, targets, and similar text in bulk according to rules

Handling and Correcting Problem Annotations​

When a collection action, the basic data, or a pose is abnormal, mark the corresponding annotation as invalid.

  1. In the annotation tree of the annotation interface, select an annotation node.
  2. Select a problem type in the pop-up menu.
  3. The marked node appears with a red strikethrough; selecting the same problem type again clears the mark.
  4. Save the annotation.

The available problem types are listed below.

Problem typeAvailable scenarios
Invalid collection dataAnnotation and check
Invalid basic dataAnnotation and check
EE result anomalyAnnotation and check
Abnormal pose dataAnnotation and check
Incorrect semantic descriptionCheck mode
Annotation node does not match the actionCheck mode

The marked results are collected on the Problem annotation tab of the task details page.

The procedure for correcting a problem annotation is as follows.

  1. Open the Problem annotation tab on the task details page and check the problem type and the correction time.
  2. Click the dataset name in that row to open the dataset in check mode.
  3. Correct the annotation content or clear the problem mark on the annotation.
  4. After saving, the correction time for that row updates accordingly.

When the correction time shows Uncorrected, the problem annotation has not been handled yet.

Submitting for Check and Follow-up​

  1. After annotating every dataset in the task, go to Current Workflow in the sidebar of the task details page.
  2. Select Next Step: Pending Check; the task status becomes Pending Check and the task passes to an auditor for review.
  3. Once an auditor marks the task Qualified, the annotator can no longer change its status.
  4. If an auditor marks the task Unqualified, it enters the Unqualified status, and the sidebar shows the Rejection Reason entered during review.
  5. After correcting the annotations according to the reason, the annotator selects Next Step: Pending Check to resubmit.

Verification​

CheckExpected result
Annotation savedThe dataset Annotation Count increases and a new node appears in the annotation tree
Batch annotationThe datasets under the target task show the annotations, which can be checked on Annotation Statistics
Problem annotationInvalid annotations are listed on the Problem annotation tab; before handling, the correction time shows Uncorrected
Status transitionAfter submission, the task status is Pending Check
ReworkWhen the task is Unqualified, the sidebar shows the Rejection Reason

Error Handling​

SymptomPossible causeResolutionOwner
The task cannot be set to a review-related statusAn annotator cannot set Qualified or UnqualifiedUse a status the annotator can set, such as In Progress or Pending CheckThis role
A message states that the task is already Qualified and cannot be modifiedThe task status is QualifiedStop editing; to rework, an auditor sets the task to UnqualifiedAuditor
A message states that the account is not related to the taskThe task is not assigned to the annotatorContact a project manager to reassign itProject Manager
Saving stops with a connection interrupted messageThe network is temporarily unavailableThe annotation draft is stored locally; save again once the network recoversThis role / Operations
The left menu is missing AnnotationRole or module permissions are insufficientContact an administrator to check the role and the View Annotation Tasks permissionAdministrator

Create Annotation Task

Dataset list

Advanced filter on the dataset list

Advanced filter on the dataset list

Advanced filter on the dataset list

Advanced filter on the dataset list

Advanced filter on the dataset list

Advanced filter on the dataset list

PagePurpose
Annotation TasksTask state machine, task details, and batch actions
Dictionary ManagementAnnotation conventions for skills, objects, and labels
Auditor GuideReview and rejection workflow
Getting StartedCommon sign-in and navigation steps
Data QCQC rules and result definitions
PermissionsRoles and permission hierarchy