Captioning QA

Overview

The Captioning QA plugin is a model plugin that creates visual QA attachments for captioning or description review workflows.

Use this plugin when a task contains bounding boxes with descriptions or OCR text and you want compact image attachments showing each crop next to its text.

Captioning QA

Plugin Functionality

The Captioning QA plugin reads the current task annotations, generates summary images, uploads those images to S3, and attaches them back to the asset.

For each image in the asset, the plugin creates an attachment containing cropped bounding-box regions and their corresponding text. The text is taken from:

• OCR text on the bounding-box annotation, when present.

• A nested classification titled description, when OCR text is not present.

For multi-image assets, the plugin can also create a correspondence attachment. It looks for a nested classification titled corresponding id and uses that value to group related object IDs across the first two images.

If the task has no tool annotations, the plugin finishes without adding attachments.

Supported Data Types

• Image

• Multi-Image Assets

Supported Annotation Tools

• Bounding Box

• OCR text on bounding boxes

• Nested classifications named description

• Nested classifications named corresponding id for correspondence views

Plugin Configuration

By default, the plugin uploads generated QA attachments to an Ango-managed S3 location. You can provide your own storage settings in the Config JSON field.

{
  "storage_id": "",
  "bucket_name": "",
  "folder": ""
}

• "storage_id": Storage integration ID used to upload the generated QA images.

  • Example:

    • "storage_id": "storage_id_here"

• "bucket_name": Name of the S3 bucket where the generated images should be uploaded.

  • Example:

    • "bucket_name": "acme-qa-attachments"

• "folder": Folder prefix where generated images should be uploaded.

  • Example:

    • "folder": "captioning-qa/"

If you provide custom storage settings, make sure the storage integration has permission to upload files and that generated attachment URLs can be accessed by Ango Hub users who need to review them.