Face Recognition (ArcFace + OpenVINO)

Pipeline completa per face recognition con:

• ArcFace (embedding 512-D) convertito in OpenVINO
• Face detection (intel/face-detection-retail-0004) + landmarks (intel/landmarks-regression-retail-0009)
• Valutazione con ROC/AUC/EER
• Webcam demo con overlay pipeline, top-K, enrollment persistente e salvataggio dell’indice
• Supporto Windows, Linux/WSL, CPU

---

---

Face Recognition (OpenVINO + ArcFace)

Demo di riconoscimento facciale in tempo reale con OpenVINO: pipeline: Detect → Landmarks → Align → Embedding → Search (Top-K)

Requisiti

• Python 3.10–3.12
• OpenVINO, OpenCV, NumPy, scikit-learn

python -m venv myvenv
# Windows
myvenv\Scripts\activate
# Linux/Mac
source myvenv/bin/activate

pip install -r requirements.txt

Modelli & Dati

Detector: intel/face-detection-retail-0004

Landmarks: intel/landmarks-regression-retail-0009

Embedding: public/face-recognition-resnet100-arcface-onnx/FP32/*.xml

Dataset LFW allineato in data/aligned/lfw/

Allineamento dataset

Se parti da immagini grezze:

python tools/align_dataset_omz.py \
  --det-xml intel/face-detection-retail-0004/FP32/face-detection-retail-0004.xml \
  --lmk-xml intel/landmarks-regression-retail-0009/FP32/landmarks-regression-retail-0009.xml \
  --src data/raw/lfw \
  --dst data/aligned/lfw

Output: cartelle data/aligned///*.jpg 112×112.

Costruzione gallery (.npz)

python tools/build_gallery.py \
  --config configs/arcface_openvino.json \
  --root data/aligned/lfw \
  --out data/index/lfw_arcface_raw255_bgr.npz

Output:

embeddings (N×512 float32, normalizzati)

labels (N oggetti)

paths (N path)

Valutazione (AUC/EER)

python src/eval_verify_openvino.py \
  --xml public/face-recognition-resnet100-arcface-onnx/FP32/face-recognition-resnet100-arcface-onnx.xml \
  --pairs 60000 --swaprb 0 --preproc raw255 --split lfw

Stampa AUC, EER, soglia a EER, TPR@FPR target.

Demo Webcam + Enrollment Windows (PowerShell)

python demo\webcam_identify_ui.py `
  --config configs\arcface_openvino.json `
  --index data\index\lfw_arcface_raw255_bgr.npz `
  --det-xml intel\face-detection-retail-0004\FP32\face-detection-retail-0004.xml `
  --lmk-xml intel\landmarks-regression-retail-0009\FP32\landmarks-regression-retail-0009.xml `
  --tau 0.30 --topk 5 --source 0 `
  --enroll-dir data\enroll --enroll-name Matteo `
  --index-save data\index\lfw_plus_me.npz

Linux

python demo/webcam_identify_ui.py \
  --config configs/arcface_openvino.json \
  --index data/index/lfw_arcface_raw255_bgr.npz \
  --det-xml intel/face-detection-retail-0004/FP32/face-detection-retail-0004.xml \
  --lmk-xml intel/landmarks-regression-retail-0009/FP32/landmarks-regression-retail-0009.xml \
  --tau 0.30 --topk 5 --source 0 \
  --enroll-dir data/enroll --enroll-name Matteo \
  --index-save data/index/lfw_plus_me.npz

Tasti utili

e → Enroll volto migliore del frame (salva crop e embedding, marca index “dirty”)

s → Save manuale dell’indice (--index-save o sovrascrive l’indice caricato)

q → Quit (se autosave abilitato, salva prima di uscire)

'+' / '-' → cambia soglia τ

a → abilita/disabilita align 5-point

p → screenshot finestra