Annotate data (Human-in-the-Loop)

更新时间:
复制 MD 格式

The Embodied Intelligence Platform integrates Label Studio for data annotation. You can create annotation tasks with built-in or custom templates, configure automated pre-annotation operators, and export results in multiple formats.

Annotation templates

The platform ships with seven built-in Label Studio templates for common annotation tasks:

Template

Data type

Image classification

Image

Image annotation

Image

Object detection (bounding box)

Image

General image template

Image

Video classification

Video

Video time segment annotation

Video

General video template

Video

Browse built-in templates

  1. In the left navigation pane, choose Data Management > Labeling jobs.

  2. Click the Labeling templates tab.

  3. Click Details next to a template to inspect its XML configuration.

Create a custom template

When the built-in templates do not cover your use case, you can define a custom annotation interface.

  1. On the Labeling templates tab, click Create custom template.

  2. Author the annotation interface in XML using the Label Studio template editor.

  3. Verify the annotation interactions in the preview area, then save.

Note

Templates use Label Studio XML syntax. For the full tag reference, see the Label Studio documentation.

Create an annotation task

  1. On the Labeling tasks tab, click Create labeling task.

  2. Set the following parameters:

    Parameter

    Required

    Description

    Task name

    Required

    A descriptive name for the annotation task.

    Labeling type

    Required

    Select Video or Image. The template list updates to match the selected type.

    Labeling template

    Required

    Choose a template that matches the selected data type. Video offers 3 templates; Image offers 4.

    Annotator

    Required

    Assign one or more annotators. You can reassign annotators after the task is created.

    Automatic labeling

    Optional

    Enable this to have an operator pre-annotate the data before human review. You must configure the operator first. For details, see Automated pre-annotation.

    Data source

    Required

    The dataset to annotate. For details on preparing datasets, see Collect and manage data.

    Expand Advanced settings to adjust the frame extraction rate for video tasks. The default is 30 fps.

  3. Click OK.

Templates by data type

Data type

Available templates

Video

Video classification, Video time segment annotation, General video template

Image

Image classification, Image annotation, Object detection (bounding box), General image template

Manage annotation tasks

The Labeling tasks tab lists all tasks with the following information:

Column

Description

Task name

The task identifier. Click the name to open task details and access the Label Studio annotation page.

Labeling template

The annotation template bound to this task.

Status

Current task status: pending, in progress, or completed.

Labeling method

Manual or automated (with pre-annotation).

Progress

The percentage of items annotated.

You can perform the following actions on each task:

  • Run pre-labeling: Invoke the configured operator to pre-annotate all unlabeled items.

  • Mark as complete: Accept the pre-annotation results as final annotations.

  • View labeling results: Review completed annotations.

  • Open Label Studio: Launch the Label Studio interface to annotate or correct items manually.

Automated pre-annotation

The platform provides six built-in operators powered by the Model Studio OpenAI-compatible API. These operators pre-annotate data so that human annotators only need to review and correct the results.

Operator

Task type

Object detection

Bounding box annotation for images and videos

Segmentation

Semantic segmentation for images and videos

Temporal segmentation

Time-segment annotation for videos

Object tracking

Multi-object tracking across video frames

Classification

Category labeling for images and videos

Keypoint

Keypoint annotation for pose estimation

Set up an operator

Each operator must be configured before you can use it in an annotation task.

  1. On the Labeling operators tab, click Configure next to the target operator.

  2. Enter your Model Studio API key and select a model.

  3. Click OK.

Once configured, turn on Automatic labeling when creating a task and select the operator. The system pre-annotates the data, then annotators review and correct the results in Label Studio.

Export annotation results

Completed annotations appear on the Labeled Datasets tab of the Datasets page. You can export results in JSON format directly, or use the Label Studio export feature for additional formats such as COCO, Pascal VOC, and YOLO.

Related documentation