Good Recognition Score Threshold

[ceyxasm]

Issue Description
Hey @vladmandic I am using faceres and faceres-deep of human to calculate similarity scores between two faces. (Calculating embeddings, then their Euclidean distance, followed by their normalization).

for i, entry in enumerate(data):
    embedding_i = np.array(entry["result"]["face"][0]["embedding"])
    similarities = []

    # Compare with every other entry
    for j, other_entry in enumerate(data):
        if i != j:  # Skip comparing the entry with itself
            embedding_j = np.array(other_entry["result"]["face"][0]["embedding"])
            score = euclidean_distance(embedding_i, embedding_j)
            _min = 0
            _max = 1
            norm_score = null
            if (score == 0): 
                norm_score = 1
            else:        
                root = score**0.5
                norm = (1 - (root / 100) - _min) / (_max - _min)
                clamp = max(min(norm, 1), 0)
                norm_score = clamp
            similarities.append({"index": j, "score": norm_score})
    # Sort similarities by score in descending order and get the top 10
    top_similar = sorted(similarities, key=lambda x: x["score"], reverse=True)[:7]

My question is what similarity score is considered a good score for threshold. In my use case, I am capturing faces on the fly and matching them with faces in my archive. But I cannot just take the most similar face because what if the face captured is not even in the archive and hence a minimum threshold is required.
Moreover I ran a few tests with images of politicians and and even dissimilar faces have a very high score.

    {
        "filename": "../recog/images/kamala-01.png",
        "top_similar_faces": [
            {
                "filename": "../recog/images/kamala-02.png",
                "score": 0.9741146265169502
            },
            {
                "filename": "../recog/images/kamala-03.png",
                "score": 0.973418896214827
            },
            {
                "filename": "../recog/images/kamala-04.png",
                "score": 0.9723309845538807
            },
            {
                "filename": "../recog/images/kamala-06.png",
                "score": 0.9723104988689575
            },
            {
                "filename": "../recog/images/kamala-05.png",
                "score": 0.970410842672599
            },
            {
                "filename": "../recog/images/trump-04.png",
                "score": 0.9687091084578349
            },
            {
                "filename": "../recog/images/trump-01.png",
                "score": 0.9685781702956071
            }
        ]
    },

Is this expected or am I doing something wrong??
Images used for the experimentation can be found here

[ceyxasm]

Also an additional question about the faceres model. In the credits you have mentioned HSE-faceres repo. There they provide mobilenet and resent models.
Is the faceres (TFJS) model used in human same as their vgg-mobilenet model??