girenit-Gemstone-Person-Object-Detector-Small

Gemstone Person & Object Detector Small

A 512 px, 80-class COCO detector adapted for the T3 Gemstone O1 / TI AM67A deployment path. This Small variant prioritizes higher detection accuracy. The primary model detects the COCO person class and objects. The repository also includes an optional local YuNet + SFace companion that recognizes faces only after an operator explicitly enrolls that identity; everyone else remains unknown.

Evidence

Training is a deterministic 512 px resolution-adaptation continuation from yolov8s.pt on 35% of COCO 2017 train for 12 epochs. Metrics use the full COCO val2017 split. See evidence.json for the complete manifest.

Files

• model.pt: Ultralytics/PyTorch checkpoint
• model.torchscript: static batch-1, 512 px TorchScript
• model.onnx: static batch-1, 512 px ONNX opset 12 without embedded NMS
• face_identity.py: consent-based local enrollment and face matching
• download_face_models.py: pinned, checksum-verified OpenCV model downloader

T3 Gemstone deployment status

The T3 Gemstone O1 has a TI AM67A, dual accelerators totaling 4 TOPS, and 4 GB RAM. The ONNX graph is prepared for the TI TIDL import path. Physical-board TIDL compilation, latency, accelerator offload percentage, and peak RAM are still pending and must not be inferred from the A100 build-host benchmark.

Quick start

from ultralytics import YOLO

model = YOLO("model.pt")
results = model("camera.jpg", imgsz=512)

For enrolled face identification, see FACE_IDENTITY.md. Face embeddings are biometric data. They are intentionally not bundled, uploaded, or sent to a remote API. This prototype has no liveness check and must not be the only signal used for authentication or consequential decisions.

Türkçe özet

Bu model T3 Gemstone O1 üzerinde kişi ve genel nesne algılar. İsteğe bağlı yüz modülü, yalnızca cihazdaki galeriye açıkça kaydedilmiş kişileri eşleştirir; diğer yüzler unknown kalır. Biyometrik galeri buluta yüklenmez. Fiziksel kart TIDL testleri tamamlanmadan gerçek zamanlı FPS iddiasında bulunulmaz.

License and attribution

Ultralytics YOLO and these derivative weights are released under AGPL-3.0. COCO annotations are CC BY 4.0; individual images retain their respective licenses. Training and packaging: girenit / Goktug Düşünen, 2026.