PCA of Eurasian male facial variation

[Siginulfo]

This PCA is the result of months of data mining and processing. I have gathered 4991 pictures of male faces representing 88 nationalities from Eurasia and North Africa, which I vectorized by the extraction and encoding of 128 facial landmarks through a script. Pigmentation was excluded from the analysis altogether.



Salient points:

• There is a great deal of overlapping between populations, especially in the case of Central and Northern Europeans
• Most of Eurasian male facial variation is best represented by a cline from "pure" Europids to MENA/South Asians, and from the latter group to "pure" Mongolids, mirroring what we are used to seeing in autosomal PCAs; there are a few outliers, mostly resulting from admixture between Europids and Mongolids (Khanty, Turkmen, Uzbek)

In the near future, I'm developing an application for ethnicity assignation for personal use, with perhaps even nMonte and Oracle modes.

[michal3141]

That's a cool project! Is your dataset publicly available? I guess these are grayscale images?

[Siginulfo]

Is your dataset publicly available?

Nope, I gathered the images "by hand" from the web.

[Creoda]

Was not being able to see anything part of your plan?

[Siginulfo]

Was not being able to see anything part of your plan?

I can see most populations pretty well. Anyway, there's not much one can do to make it easier to visualize a PCA with ~5000 data points.

[michal3141]

Nope, I gathered the images "by hand" from the web.

OK. Could you post some of the Polish samples you used? Are these ordinary people or celebrities/athletes?

[Creoda]

I can see most populations pretty well. Anyway, there's not much one can do to make it easier to visualize a PCA with ~5000 data points.

Maybe zoom in a bit and turn the colours off.

[Rico33]

I think that image will cause epilepsy to some people.

[Bellbeaking]

Nice work OP.

Can we get a zoom in on the packed European part?