Is dlib_face_recognition_resnet_model_v1 better than HSE_FaceRec_tf?

[kekoleg]

I am compairing face-recognition.js (is using dlib_face_recognition_resnet_model_v1) and human. I have a set of 130 faces (different people) and another one that belongs to the person in this set and is used as a search input. Let's name the person we are looking for in the set x. I want to have x as the closest match to the input. Both face-recognition.js and human successfully found the searched person as the closest, but I've noticed one thing. The distance to the next person from the match is smaller in the human search result compared to the face-recognition.js search result. Does this mean that face-recognition.js was more confident in distinguishing faces and human will produce more false positives as the database grows significantly?

human

Person	Distance	Match
x	7.644384539662718	✔️
	10.066655598143836
	10.356641958367035
	10.445426443843736
	10.521632118242902

face-recognition.js

Person	Distance	Match
x	7.644384525550511	✔️
	15.685646888231423
	16.320001040515428
	16.522694225317679
	17.079407024941559

(sorted by smallest distance)

To compare the search results from both libraries I had to multiply every face-recognition.js distance by 27.3558887. Am I doing something wrong? Am I jumping to the wrong conclusions?

[vladmandic]

sounds about right. but note that its a difficult topic.

• different face recognition models are trained on different face crop factors, so having wrong crop factor may show one model better than the other and changing crop factor may change it upside-down
• human supports several different face recognition models built-in

[kekoleg]

@vladmandic do you have any plans to support dlib_face_recognition_resnet_model_v1 in human? It's a really good high quality model, but human is generally much better than face-recognition.js in terms of resource consumption. Combining these things will be an excellent solution.

[vladmandic]

resnet models are generally much larger and not really suitable for web usage, so its not primary target for human.

[shurshilov]

How often does the person who created something not even understand how to use it. You are wrong, even if the model is 50 megabytes, it will be used, for example, in kiosk mode where the model is loaded once, but greater accuracy will improve everything

[vladmandic]

you're entitled to your opinion without the need to be insulting.

[shurshilov]

I didn't insult you, I just see very often when the creators of a library get into a dead end without understanding how to develop it, because don't know the application part. This is useful information for you and your product. But you're probably used to listening to only positive reviews... In general, by the way, false positives are also many. But this is another stage that has to be bypassed by checking the alignment of the eye position and so on. The recognition itself seems to be in error. Also anti-spoofing its just check blink eyes;)

[vladmandic]

You're assuming far too much about "usual developer" and never bothered to ask. That's fine, assume away. I'm not in habbit of pointless arguing, I'm out of this thread.