CompreFace
Face Recognition Similarity Threshold
The result of CompreFace face recognition and face verification services is a similarity between faces. Even if you upload the faces of two different people, you still receive the result, but the similarity is low. Therefore, the user must determine for himself whether this is the same person or not using similarity. The level of similarity the user accepts, as big enough, we call similarity threshold.
How to choose the face similarity threshold
No Face Recognition Service has 100% accuracy, so there always appear errors in recognition. If a user chooses too low a threshold, then some unknown faces are recognized as known. If a user chooses too high a threshold, then some known faces are recognized as unknown. CompreFace calculates similarity so that most correct guesses have a threshold of more than 0.5, and the most incorrect guesses have a threshold of less than 0.5. Still, we recommend for high-security systems set the threshold more than 0.5. This is the distribution of similarities for a custom dataset of 50,000 faces for the FaceNet model (blue - is incorrect guesses, red is correct):
