Detect Objects by Text Prompt (Ultralytics YOLO World)

How it fits

Typical backends

Caveats

• I/O contractThe output is a oneof resolved at deploy time by the connected sink, and every shape is a projection of one per-detection record so the shapes agree detection-for-detection: [BoundingBox] is box geometry plus class, [Rectangle<Double>] is box geometry only, [DetectedClass] is class id plus confidence only, [(BoundingBox, DetectedClass)] pairs each box with its class, UInt64 is the detection count, and Image is the source frame with boxes drawn.
• Hard constraintThe bound weights file must resolve to exactly one YOLO-World checkpoint, and the text class list must hold at least one non-empty name; either being unmet aborts startup. Only the named categories are detected.
• Parameter interactionAn emitted class id is the zero-based index into the configured class list, and the per-class confidence overrides are keyed by that same prompt index; a class without an override uses the global confidence floor. The IoU bound caps NMS overlap on every arm.
• AccuracyThe default confidence floor is deliberately low because open-vocabulary scoring sits below fixed-vocabulary detection; raise it to suppress weak matches. The inference image size should match the training-time input size; a mismatch silently degrades accuracy.
• LimitationThis is a detection-only model with no instance-mask output arm; Ultralytics ships no YOLO-World-seg head. For open-vocabulary masks use a segmentation component instead.
• FallbackA frame with no match for any prompted name yields the safe empty value for the resolved arm (an empty list, a zero count, or a blank overlay frame). A device setting starting with `cuda` silently falls back to CPU when CUDA is unavailable.
• State lifetimeAll configuration keys, the bound weights, and the embedded text vocabulary are captured once at startup; runtime changes have no effect and require a redeploy.