Attribute annotation

Attribute annotation adds qualifiers to an existing label. Once an object is annotated with a class and a shape, a box or mask, attributes capture its properties: a car's color, a traffic light's state (red, yellow, green), a pedestrian's pose, or whether an object is occluded or truncated.

Attributes can be categorical (color), boolean (is_occluded), or numeric, and they are defined in the label schema or ontology so they are applied consistently. They let a dataset express fine distinctions without multiplying classes: instead of "red car," "blue car," and "white car" as separate classes, you have one "car" class with a color attribute.

Attributes are defined per class in the schema, then set by annotators or models on each label. In FiftyOne, labels carry attributes alongside their class and geometry, so you can filter and slice a dataset by attribute, all occluded pedestrians, all red lights, and evaluate model performance within each.

Attributes are how a dataset captures the nuance that drives real model behavior, and they are the cleanest alternative to class explosion. Attributes are also the most under-audited part of a label, because review attention goes to the box or mask, so the attribute fields are where silent, systematic errors accumulate, a default value left unchanged, an occlusion flag applied inconsistently. Yet attributes often carry the safety-critical signal, a traffic light's state matters more than the box around it, so slicing evaluation by attribute frequently reveals failures the overall metric hides.

Labeling descriptive properties on an object, like color, state, or occlusion, in addition to its class.

Attributes keep the schema clean and avoid a combinatorial explosion of near-duplicate classes.

In the label schema or ontology, so they are applied consistently.

Label schema, Annotation ontology, Bounding box, Annotation guidelines, Annotation.