Overview

The Labeling feature allows you to label your data to prepare for model training. Currently, classification and object detection types of labeling jobs are supported.

Auto labeling leverages advanced models to produce machine-generated labels automatically. Manual labeling can use a human expert to label your data or review and correct the autolabel-produced labels.

Refer to the video on Overview of the Labeling Feature.

Key features

• Auto labeling uses advanced AI algorithms to automatically identify and label objects within images or video frames. It aims to significantly speed up the data labeling by automating much of the work. Auto label process combines language and image sample inputs to create labels.

• Manual labeling is a service offering in which a human expert is assigned to review existing labels or add new labels to the dataset.

Overall, the auto labeling and manual labeling features aim to provide a hybrid solution utilizing the speed of auto labeling with result improvement from human experts.

This process involves:

• Creating a label spec using text descriptions and image examples. Advanced large language models are used to refine and disambiguate the text input.

• Creating auto-label jobs and manual label jobs.

• Review the label job results and bucket them into accept or reject categories.

• Creating result sets from label jobs and exporting them to the catalog or downloading the labels for downstream tasks like training. A typical usage mode is to export accepted labels to the catalog and feed the rejected labels into a manual labeling job for further refinement and correction.