Detect Objects (RF DETR)

How it fits

Typical backends

Caveats

• Hard constraintThe bound weights directory must contain exactly one `.pth` checkpoint, and variant must match the trained backbone; a mismatch fails checkpoint loading with a state-dict error at startup.
• 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, UInt64 is the detection count, and Image is the source frame with boxes drawn. RF-DETR is a set predictor that de-duplicates internally, so no NMS runs. This is a box-only detector, so no instance-mask shape is offered.
• Parameter interactionthreshold is the base confidence floor, class_thresholds overrides it per class id, and object_classes is an allow-list applied after inference (an empty list keeps every class), so the model still runs the full forward over each frame regardless of the filter.
• AccuracyThe class id mapping is checkpoint-dependent, so downstream logic that branches on a class id will misfire if the bound checkpoint changes.
• State lifetimeThe bound checkpoint, variant, and all configuration keys are captured once at startup; replacing any of them requires a redeploy.
• Fallbackdevice starting with `cuda` silently falls back to CPU when CUDA is unavailable.