Threshold of trained models

[mohamedamara7]

How I can know the best threshold of your trained models for face verification? Is there even a pseudocode for reading two faces using cv2 and compare their similarity?

[leondgarse]

• Most time I'm using 0.6 as threshold.

• For reading image and compare, you can refer to colab efficientnetV2_basic_test.ipynb Keras insightface test section. Technically, Just following the steps face detection in image --> face align --> normalize to [-1, 1] --> input to model and get embedding output --> l2 normalize embedding --> calculate cosine distance is enough. Processes in video_test.py#L90 is following this.

  iaa = np.zeros([1, 112, 112, 3]) # Face image 1
  ibb = np.ones([1, 112, 112, 3]) # Face image 2
  
  mm = keras.models.load_model('model_path.h5') # load pretrained basic_model
  eaa = mm(iaa)  # face 1 embedding
  ebb = mm(ibb) # face 2 embedding
  print(eaa.shape, ebb.shape)
  # (1, 256) (1, 256)
  
  from sklearn.preprocessing import normalize
  eea = normalize(eaa) # normalize
  eeb = normalize(ebb)
  print((eea ** 2).sum(), (eeb ** 2).sum())
  # 1.0 1.0
  
  print(np.dot(eea, eeb.T)) # cos similarity
  # [[0.38431713]]