High percentage in similarity

[Arab04]

1. When we compare two different faces it returns high similarity in face verification REST API what is the reason for that?
2. How many points does it uses when it makes face verification?
3. I saw that there are 5 main points and does that mean that it puts only 5 points to the face when it verifies user?
4. what type of algorithms does it use when it verifies user face?

I really need answers to all this questions, I would be very glad if someone could give me explanation to my questions

Thanks in advance

[pospielov]

Sorry for the late reply.

1. There could be lots of reasons:

• The bad picture quality or extreme lighting, for example, severe backlighting
• Face doesn't look to the camera (extreme angle)
• The bad accuracy of the model - in this case, I can recommend using one of the custom builds (https://github.com/exadel-inc/CompreFace/tree/master/custom-builds) or training your own model on your faces.

2. Default version of CompreFace scales the image of the face to 160*160 px and sends it to a neural network. A neural network produces the embedding of 512 numbers. Embedding is used to verify the faces.
3. Landmarks are internally used to alight the face, not to verify them.
4. Euclidian distance between embeddings.

[pospielov]

1. There could be lots of reasons:

• The bad picture quality or extreme lighting, for example, severe backlighting
• Face doesn't look to the camera (extreme angle)
• The bad accuracy of the model - in this case, I can recommend using one of the custom builds (https://github.com/exadel-inc/CompreFace/tree/master/custom-builds) or training your own model on your faces.

2. Default version of CompreFace scales the image of the face to 160*160 px and sends it to a neural network. A neural network produces the embedding of 512 numbers. Embedding is used to verify the faces.
3. Landmarks are internally used to alight the face, not to verify them.
4. Euclidian distance between embeddings.

[Arab04]

Hi, thank you for your reply, I would like to ask about the server that we should use for deploying compreFace, could you please tell: 1. Is it enough for server with capability 4x16 without GPU, to achieve 1 second for verifying face 2. If not , could you please tell what kind of server should be used for compreFace with criteria like ram, storage, cpu architecture and etc. thank you for your help in advance

…

On Thu, Sep 29, 2022 at 2:44 PM Pospielov Serhii ***@***.***> wrote: 1. There could be lots of reasons: - The bad picture quality or extreme lighting, for example, severe backlighting - Face doesn't look to the camera (extreme angle) - The bad accuracy of the model - in this case, I can recommend using one of the custom builds ( https://github.com/exadel-inc/CompreFace/tree/master/custom-builds) or training your own model on your faces. 1. Default version of CompreFace scales the image of the face to 160*160 px and sends it to a neural network. A neural network produces the embedding of 512 numbers. Embedding is used to verify the faces. 2. Landmarks are internally used to alight the face, not to verify them. 3. Euclidian distance between embeddings. — Reply to this email directly, view it on GitHub <#902 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQFOIOIP566RE56IHVRQEZ3WAVQHJANCNFSM6AAAAAAQRWRTXA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[pospielov]

A server with 4 cores and 16Gb of RAM is totally enough to achieve 1 second for verifying face

[Arab04]

Thank you ср, 23 нояб. 2022 г., 15:24 Pospielov Serhii ***@***.***>:

…

A server with 4 cores and 16Gb of RAM is totally enough to achieve 1 second for verifying face — Reply to this email directly, view it on GitHub <#902 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQFOIOLB6IG7C7ELRSHOS33WJXWF7ANCNFSM6AAAAAAQRWRTXA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[dingyaguang117]

For Asians, 1.1.0-arcface-r100 works fine for me.