git clone https://github.com/Daedaluz/rknn-docker.git
cd rknn-docker
sudo docker build -t rknn-lite .
docker run -it --rm \
    -v $(pwd):/workspace \
    rknn-lite \
    bash
cd wirkspace
python3 convert_models.py


poi
python3 test_rknn.py
risultato
Runtime init OK (CPU mode)

SCRFD (Face Detector)
https://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/det_2.5g.onnx
opz
https://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/det_500m.onnx
https://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/det_10g.onnx

ArcFace
leggero
https://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/w600k_mbf.onnx
pesante e piu accurato
https://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/w600k_r50.onnx


SCRFD (Face Detector) – ONNX diretto

Dalla release del progetto face‑reidentification che include i modelli SCRFD:

SCRFD 2.5G

https://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/det_2.5g.onnx

(Opzionali)

SCRFD 500Mhttps://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/det_500m.onnx

SCRFD 10Ghttps://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/det_10g.onnx

🔹 ArcFace (Face Recognition) – ONNX diretto

Dalla stessa release, modelli ArcFace in ONNX:

ArcFace MobileFace (veloce, leggero)

https://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/w600k_mbf.onnx

ArcFace ResNet‑50 (più pesante, più accurato)

https://github.com/yakhyo/face-reidentification/releases/download/v0.0.1/w600k_r50.onnx

📌 Consiglio tecnico per RK3588

Per Orange Pi 5 Plus:

SCRFD 2.5G → miglior compromesso velocità/precisione

ArcFace MobileFace (w600k_mbf.onnx) → più leggero, conversione RKNN più stabile