Issue with Inter-eye distance (facial recognition)

[Huyanyinglei]

TL;DR

face alignment and inter-eye distance

I have the following problem. I’m a beginner of facial landmark detection or so-called face alignment. I see others use Inter-eye distance when they’re doing face alignment. E.g., this: https://github.com/patrikhuber/superviseddescent/issues/56, uses IED for normalization. I wonder what role the IED plays in face alignment methods in general, why and when we need to do it and how we can do it.
May I ask if there’s some good, easy-to-read articles that can help beginners like me understand this question? After some google search I didn’t find much about it. Thanx in advance.

 

[sysprog]

You can upload a face image to morphthing.com, and click on points of reference, and then check out the result as the image is combined with another that's in their library, or with another that you've uploaded and selected points for -- it's fun even to just play around with A+B combinations from their library, but doing the point selection is more instructive in giving you some idea of what must be happening in the processes -- e.g. watch how wonky it looks if you deliberately pick off-target points.

 

[sysprog]

You can find some accessible MIT articles here: https://www.technologyreview.com/artificial-intelligence/face-recognition/

Facial recognition is not easy -- one encounters complex versions of Sayre's Paradox -- but take heart -- you phone has much more processing and storage capacity than had the IBM machines that were available to Prof. Sayre in 1973.

Here's a link for some open source facial recognition software: https://cmusatyalab.github.io/openface/

If you read some of the documentation there, you'll gain a deeper understanding.