Table of Contents
Updated by Alex Cota
By September 30, 2020, Labelbox customers labeling images, video, and text data will be expected to use the new editor for all projects. Only projects for labeling geospatial data will be supported in the legacy editor. This document is intended to provide our customers with the general guidelines and resources for migrating their projects, members, and ontologies from the legacy editor to the new editor in Labelbox. It also includes a side-by-side comparison of the UI differences.
There are a few major limitations of the legacy editor data model.
- First, there is no concept of global ontologies since all Label information is contained at the project level. This makes it difficult for customers to reuse annotation classes from project to project.
- Secondly, sometimes customers experience issues working with individual annotations in a Label because individual instances of an annotation class are not stored as unique annotations in the data model. In the end, these limitations end up surfacing in the app as a limited user interface and limited functionality within the legacy editor.
- Lastly, the legacy labeling frontend options are supported by a persistent data model.
After spending time learning how the limitations of the legacy editor have impacted our customers’ ability to collaboratively create, review, and analyze their training data, we created a new and improved editor with a more granular and persistent data model to support our customers’ needs. In the new editor, annotation classes are organized into reusable and persistent ontologies for use across multiple projects. Also, the structure of an ontology is enforced at the data model level, which means improperly formed tools or classes won’t be accepted. This more granular and persistent data model reliably supports the create, read, edit, and delete actions on individual annotations in a Label and enables Labelbox to create more robust labeling tools, QA tools, and dashboards to organize and analyze your training data.
For a visualization, see Legacy vs. new editor ontology.
The table below compares the features and capabilities of the new editor to the legacy editor. All of the features in the legacy editor (with the exception of Tiled Imagery and dark mode) are supported in the new editor.
Contact our support team for assistance with the following migration steps for each of your projects:
- Make a new project in the new editor.
- Grab the ontology from the old editor and attach it to the new project.
- Move existing members to the project in the new editor.
- Grab the project’s dataset and add it to the new project.
Data import format
There is no difference between importing image data in the legacy vs the new editor. Note that only the new editor natively supports video and text imports.
See our Data import overview for import instructions.
Label export format
There are some key differences between the legacy and new editor export formats. The new export format reflects the new data model that stores each annotation class instance as an individual, schematized annotation.
To see a side-by-side comparison of the legacy vs new Label export formats for each annotation type, see our docs on Label export format (legacy vs new editor).
Legacy ontologies are backwards compatible in the new editor. You may pass an ontology using the format from the legacy editor to the new editor and Labelbox will automatically schematize your ontology for you.
To see a side-by-side comparison of the ontology formats from the legacy editor to the new editor, see our docs on Ontology format (legacy vs new editor)
Updating the ontology
In the new editor, to make a change to your ontology, you’ll need to keep the existing
featureSchemaId when you make your modifications to avoid creating a duplicate schematized class. For example, in the legacy editor, all that is needed to make a change to the ontology is to change the string of the class. However, in the new editor, you’ll need to make sure that whatever modifications you make, the
featureSchemaId remains the same to avoid creating a duplicate class in your ontology. For more information see Legacy vs new editor ontology.
Here are the key changes that may impact your API integrations with Labelbox.
Programmatic label creation
updateLabel mutations are not supported in the new editor. At this time, Labelbox does not support programmatic label creation in the new editor due to a conflict in data types since the new editor uses a completely different data model.
The import format and the process for bulk importing predictions in the new editor is completely different. Any interactions with legacy
predictionModels are not supported in the new editor.
For instructions on how to programmatically import predictions in the new editor, see our docs on Model-assisted labeling.
Most of the hotkeys (shortcuts) from the legacy editor to the new editor are the same, with some minor changes. Though we added a lot more shortcuts to the list. For a side-by-side comparison, see Legacy vs new editor shortcuts.
- Ontology search, sharing, and cloning not supported.
- Ontologies can only have vector annotations. Not possible to have vector annotations (Rectangle, Polygon, Line, Point) and mask annotations in the same Label.
- Programmatic configuration in Configure editor step.
- Admins can search for an ontology to reference (share) or clone an existing ontology.
- Can have all annotation types (Mask, Bounding box, Polyline, Point, Polyline, Radio, Checklist, Dropdown, Text) in one Label.
- No programmatic configuration in app but you can create an ontology programmatically using our Python SDK. To see an example, see our end-to-end project script.
- In the new editor, “Rectangle” has been renamed to “Bounding box” and “Line” has been renamed to “Polyline”.
- No display of each individual instance of an annotation class.
- Mask templates are not supported.
- No way to search large ontologies.
- Drag bounding boxes not supported.
- Labelers can create unique instances of an annotation class and see them listed in the Objects panel. They can also copy/paste or duplicate annotations from the objects panel to create more instances.
- When you delete one or more labels, you have the option to create templates out of the existing masks.
- Has a search bar where labelers can more easily find the class they are looking for.
- Drag bounding boxes supported.
Although the open review and queue-based review have no significant UI changes, the new editor has more robust calculations for Benchmarks and Consensus. Also, the new editor tracks label time and review time separately. You can still mark a Label as a Benchmark by selecting a data row from the Activity table and clicking the star icon at the top.