I've got to say..

[atv2016]

I was feeling disheartened by the guesses DeepStack took, it seems to take an almost willy nilly approach to detection (and seems to accept a wider acceptance of what a face constitutes). I was about to throw it away as not useful/ready for primetime.

As i couldn't run Compreface on my intel celeron minipc, i booted it up on my Mac (which does have a genuine intel). What a shocking difference. Everything, without fail, is 98-100 percent correct. And this from 5 m away even, in all kinds of angles. Compreface seems to use a much more advanced and scientific approach, whereas DeepStack feels like a amateur opencv project.

This i can actually use for authentication or intruder detection. I can imagine adding this to a composite presence tracker. I'm going to have a lot of fun with this for sure.

There's one thing that seemed to pop up, for some reason it sees some spiderweb as an trained person. Not sure how to resolve that one. I guess you can't untrain an image? What do i do, put it in a false positive folder maybe?

Thanks @pospielov.

[pospielov]

There are different options:

1. As you said, put the spiderweb photo to the "false positive" subject
2. Increase det_prob_threshold parameter, so it's not recognized spiderweb as a face
3. If the similarity with the trained person is low, then just have a custom logic that counts it as "false positive"
4. Try one of the custom builds. Probably face detection will work differently: https://github.com/exadel-inc/CompreFace/tree/master/custom-builds