Open-Vocabulary Detection

Prompts are encoded by the OWLv2 text encoder (CLIP ViT-B/16) and cached by the edge node as a frozenset. Cached embeddings are reused across all subsequent frames — the text encoder runs only once per unique prompt set. The visual backbone processes every frame regardless.

After adding or removing prompts, click Push to Edge to sync via MQTT aiips/…/detection/custom-class/sync (QoS 1). The edge node acknowledges the update within the next heartbeat cycle (~5 s).

Detections are returned as { class_id: -1, class_name: <prompt>, detector: "owlv2" } and flow through the same alert pipeline as RF-DETR detections. The configured threat contribution is added to the Bayesian score at Layer 6.