Model zoo

Primary (PGIE) classroom detector for the Karterschool domain — detects and directly classifies people as adult vs child in one pass (2 classes). Preferred primary when a deployment needs the adult/child split at detection time: teacher:student ratio counting (CNT08/CNT09) and adult-targeted secondary models (karter-phone, karter-par-dinov2 run on its adult crops). Alternative primary: karter-yolo11 (person-only) + SGIE age split via karter-par-dinov2.

Secondary (SGIE) person-attribute classifier for Karterschool — DINOv2 backbone, sigmoid multi-label over 5 attributes (not_using_phone, using_phone, unknown, child, adult), run on person crops from the primary detector. This is the model that writes the `<label>_scores` attributes rules read: using_phone_scores feeds PAR02/PAR03 phone-usage rules; adult_scores / child_scores feed CNT08/CNT09 when the primary is person-only (karter-yolo11). Classifier only — it does not produce boxes.

Secondary (SGIE) phone detector — runs on adult crops produced by the child/adult PGIE and localizes a visible phone (1 class). Used to confirm phone presence with an actual bbox where the attribute-score route (karter-par-dinov2 using_phone_scores) is not confident enough on its own. Only sees crops of already-detected adults — it is not a full-frame phone detector.

Primary (PGIE) person detector for the Karterschool domain — YOLOv11, single class (person), 640x640. Use when the deployment only needs person boxes and delegates the adult/child or attribute split to an SGIE (karter-par-dinov2 emits adult/child/using_phone scores on its person crops). Higher-recall generic person detection than karter-child-adult, but no age split at detection time.

Pedestrian Attribute Recognition (PAR) classifier — sigmoid multi-label over 102 person attributes (gender, age bands, headwear incl. hat/glasses/facemask, clothing types incl. shirt_ppe/rain_coat, carried items, phone usage, ...), run as an SGIE on person crops in warehouse deployments. Emits `<attribute>_scores` per tracked person; PAR02 reads using_phone_scores from it. The broad attribute set makes it the go-to SGIE for new appearance-based person rules (dress code, carried objects, headwear) without training a dedicated model. Classifier only — no boxes.

Primary (PGIE) warehouse object detector — YOLOv8m trained on industrial warehouse footage. Emits 4 classes: person, forklift, product, pallet. This is the model behind the logical roles "warehouse_detector" and "person_detector" that the warehouse rules (MHE10, HM07, PPE01, PAR02) declare in requires.models: it provides the person boxes for loitering/PPE/phone rules and the forklift/product boxes for MHE geometry rules. Runs full-frame at 640x640; downstream SGIEs (par, PPE classifier) run on its person crops.