Question

A media company needs to label thousands of images with a proprietary taxonomy that does not match any of the built-in SageMaker Ground Truth task types (bounding box, image classification, semantic segmentation, etc.). The labeling UI must present custom instructions and a specialized annotation interface, and the raw annotations must be post-processed into a consistent label format before being written to the output manifest. Which approach lets the team implement this within SageMaker Ground Truth?

Accepted Answer

Create a custom labeling workflow using an HTML custom worker task template plus pre-annotation and post-annotation AWS Lambda functions.

Answer

Use an automated data labeling job so Ground Truth's active learning model generates the proprietary taxonomy labels without human input.

Answer

Modify the built-in image classification task type's output manifest schema in the SageMaker console to include the proprietary taxonomy fields.

Answer

Export the images to Amazon Rekognition Custom Labels, which supports arbitrary annotation UIs and writes directly to a Ground Truth manifest.