Detect Objects (Triton)

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, UInt64 is the detection count, and Image is the source frame with boxes drawn. This is a box-only detector, so no instance-mask shape is offered.
• Parameter interactionThe preprocessing keys (mean, std, color_model, change_channel_order, rescale_factor, letterbox sizing) must mirror the model's training config; any drift silently produces low-confidence results rather than an error.
• Parameter interactionbbox_x0_y0_x1_y1_order, center_to_corners, and logits decide how the raw detection tensor is parsed; a mismatch silently produces wrong coordinates or unscaled scores.
• Hard constraintnum_classes must equal the model's class count exactly — too small drops trailing classes, too large reads past the tensor end — and the output tensor names (output_score, output_label, output_detections, num_detections) must match the served signature or the result is silently empty.
• Parameter interactionobject_confidence_threshold is the base cutoff, class_confidence_thresholds overrides it per class, and classes is an allow-list where a non-empty list keeps only the listed class ids; object_nms_threshold and classes_nms_equivalence govern duplicate suppression.
• FallbackA frame that decodes empty yields the safe empty value for the resolved arm (an empty list, a zero count, or the unchanged blank frame for the overlay).
• State lifetimeAll configuration keys are captured once at startup; runtime changes have no effect and require a redeploy.
• CompatibilityA Triton inference server hosting the detection model is required; this component cannot run standalone.