Ontology management

An ontology (sometimes called a taxonomy) helps define how unstructured data should be annotated for use in model training.

Why are Ontologies important?

A clean, thoughtful ontology is critical for creating high-quality labeled data with minimal errors and inconsistencies. A good ontology defines the following objects for your labeling team:

  • What should the labeling team be labeling? What are the objects of interest or classification tasks that the labeling team should be answering about the unstructured data? The ontology should be a blueprint of the structure you hope you apply to the data in order to train your model.
  • How should objects or classifications be labeled? Your ontology should specify how you expect things to be labeled for your model. As an example, should an object be labeled with a bounding box, polygon or segmentation mask? For classification tasks, what are the options for how something should be classified
  • What additional information is helpful for your model or review team to know? Sometimes there is secondary or tertiary information that is helpful to know about the data during labeling.

For example, let's say we want to train a model on identifying dogs and cats on images. Our ontology in this example requires our labeling team to identify objects in our images. In our ontology, we should clearly specific that we want the labelers to annotate dogs and cats with bounding boxes. Furthermore, we know that some images can be blurry and that can affect our model performance, so we want annotators to answer if the image is blurry or not. This ontology is shown below as "Ontology 1"

Now let's say you have a new project to identify dogs and llamas. We can re-use some of the component of our old ontology, but need to replace the object of "cat" with a new object for "llama". This is shown in the blue box above for "Ontology 2".

Ontologies in Labelbox

Ontologies are an essential part of the Labelbox labeling platform. Each labeling project or a model in Labelbox requires an ontology.


Did this page help you?