Import annotations

Labelbox offers several ways to import annotations and attach them to assets in Labelbox. To import annotations to a labeling project, you can either import annotations as pre-labels or ground truth. You can also import annotations as pre-labels to a model run which is covered in a separate section.

If you want to import already-existing annotations in Labelbox, and you want a human to manually review and/or edit them, you can import your annotations as pre-labels. For example, you might have computer-generated predictions or simply annotations created outside of Labelbox.

By importing annotations as pre-labels on an asset, the asset will need to be labeled by a human in Labelbox. However, the imported annotations will be pre-populated in the labeling editor, to speed up human labeling.

If you want to import already-existing annotations in Labelbox, and you do not need a human to review or edit them, you can import your annotations as ground truth. For example, you might be switching to Labelbox and want to migrate annotations from another platform.

By importing annotations as ground truths on an asset, the asset will not go through any human labeling: the imported annotations become the ground truth label for the asset.

Python Annotation types vs NDJSON

When you create your annotation payload, you will need to choose between the following formats. Each comes with its own benefits.

Python Annotation Type (recommended)NDJSON
- Provides a seamless transition between third-party platforms, machine learning pipelines, and Labelbox.

- Allows you to build annotations locally with local file paths, numpy arrays, or URLs

- Easily convert Python Annotation Type format to NDJSON format to quickly import annotations to Labelbox

- For object annotations, it supports one level of nested classification for free text / radio / checklist.
- Skip formatting annotation payload in the Python Annotation Types format just to convert back to NDJSON

- Ability to create the payload in the NDJSON import format directly

- For object annotations, it supports any levels of nested classification for free text / radio / checklist.