Optional ReadonlyidOptional stable identity for picking, interaction, and host app metadata.
OptionaltrackerOptional temporal identity assigned by a tracking post-processor.
This is separate from id, which remains the annotation/picking identity.
Post-processors may update this derived field in place.
Optional ReadonlysourceOptional provenance for detections copied from a composed source.
This is renderer-neutral source identity, not product workflow state.
Optional ReadonlysourceOptional source-local detection index before composition.
Optional ReadonlyclassOptional semantic class name, such as person, vehicle, or basketball.
Optional ReadonlyconfidenceOptional confidence score from 0 to 1.
Optional ReadonlyattributesOptional host-owned annotation attributes.
Optional ReadonlyzOptional per-annotation render and pick order. Higher values are on top.
Optional ReadonlylockedOptional host editing lock. Renderers may still present locked records.
Optional ReadonlyrectOptional axis-aligned media-pixel rectangle.
Optional ReadonlypolygonOptional closed polygon in media-pixel coordinates.
Optional ReadonlypolylineOptional open path in media-pixel coordinates.
Optional ReadonlykeypointsOptional keypoints and skeleton edges in media-pixel coordinates.
Optional ReadonlymaskOptional binary mask in semantic cold-storage form.
Optional ReadonlymetadataCaller-owned metadata. The renderer does not interpret this field.
One semantic detection for a media frame.
A detection may come directly from a model or from a semantic post-processor such as tracking. It stores identity, class, confidence, geometry, mask, and caller metadata, but intentionally does not carry render styling. Styling is resolved through presentation styles so the same detections can be rendered as boxes, masks, labels, or future layers without mutating the underlying annotation data.