An annotation taxonomy is the hierarchical organization of the categories in a labeling task, arranging classes into parent-child levels, from broad groups down to specific types. It defines how labels nest, so related categories are grouped rather than listed flat.
An annotation taxonomy is the classification tree for a dataset's labels. Instead of a flat list, classes are arranged in levels: vehicle at the top, then car, truck, and bus beneath it, then sedan and SUV beneath car.
This nesting lets a model and a team reason at multiple granularities, you can evaluate at the coarse level, all vehicles, or the fine level, sedans, and annotators can label to whatever depth the data supports. A taxonomy is the hierarchy specifically, one part of a fuller ontology, which also adds attributes and non-hierarchical relationships.
Key takeaways
A taxonomy organizes label categories into a parent-child hierarchy, from broad to specific.
It enables reasoning and evaluation at multiple levels of granularity.
It is the hierarchy piece of an ontology: schema is the flat list, taxonomy adds the tree, ontology adds relationships and attributes.
What a taxonomy provides
What an annotation taxonomy provides.
What an annotation taxonomy provides.
Feature
What it provides
Levels
Broad parent classes down to specific leaves
Multi-granularity labeling
Label to the depth the evidence supports, vehicle if unsure, sedan if clear
Roll-up evaluation
Score at any level by aggregating children
A fallback for ambiguity
A parent class for cases too unclear to label precisely
How it works
The taxonomy is defined with the schema and applied during labeling, with classes nested rather than flat. In FiftyOne, hierarchical classes can be stored and analyzed, so you can evaluate and slice a dataset at coarse or fine levels and see where confusion clusters within a branch.
Why it matters
A taxonomy turns an unwieldy flat list of classes into something a team and a model can navigate. The practical payoff is graceful degradation under uncertainty. With a flat schema, an annotator unsure between sedan and coupe either guesses, adding noise, or skips, losing the sample, but with a taxonomy they can label the parent, car, capturing real signal without forcing a false precision. So a good taxonomy reduces both noise and missing labels exactly on the hard, ambiguous cases, which is where flat schemas quietly bleed quality.
Frequently asked questions
What is an annotation taxonomy?
The hierarchical organization of a dataset's label categories, from broad parents to specific children.
What is the difference between a taxonomy and an ontology?
A taxonomy is the hierarchy of classes. An ontology adds attributes and non-hierarchical relationships on top.
Why use a hierarchical taxonomy?
It enables multi-level evaluation and lets annotators label a parent class when the specific type is ambiguous.