ConceptsData

Ontologies

Versioned class and tag vocabularies that datasets and models agree on.

An ontology is a named list of the classes you draw and the tags you attach: car, pedestrian, traffic light, plus occluded or weather: rain. Labeling stages hand it to labelers, datasets record which one their annotations are written in, and models remember the one they were trained on.

Versions

Ontologies are versioned. Editing one never changes what’s already there: you edit, then publish, and the result is a new version (v2, v3…). Publishing an unchanged list creates nothing.

Every dataset is pinned to one exact ontology version — that’s the statement “these annotations are written in these words”. A pin only ever moves forward: when a labeling stage or an import adds classes, the dataset moves to the newer version, and nothing moves it back.

Class names are shown from the newest version, so renaming ped to pedestrian reaches old datasets and old predictions without relabeling or retraining.

Classes and tags

  • A class (Object in the editor) is something drawn on the image: a box, a polygon, a mask. You can restrict a class to certain shapes, or allow any.
  • A tag is attached to a whole image or to an object. It can be a plain flag, or carry a number, a text, a date, or one value from a fixed list such as day, night. A tag can apply to images, objects, or both.

Each class and tag can have a color, a one-letter hotkey for the annotation tool, and a description — a hint for labelers, which models that take text prompts also use.

Order matters. Models refer to classes by their position in the list, so reordering classes creates a new version just like adding one does.

Where ontologies come from

You rarely need to type one in from scratch:

  • Imports create or extend one from the classes found in the imported data.
  • Training records the model’s class list.
  • Labeling stages extend the ontology when labelers are allowed to add classes and tags.
  • The Create ontology node creates or edits one as part of a flow.

You can also create and edit ontologies directly on the Ontologies page in the main sidebar. Each ontology’s page shows its classes and tags, its version history, and which datasets and models use which version.

Shared ontologies

The Shared tab lists vocabularies installed with the pretrained model catalog, such as the 80 COCO classes. They’re read-only and can’t be assigned to your datasets. When you add a pretrained model to your registry, it gets its own copy in your workspace.