SFPC

School for Poetic Computation, Fall 2015

Face

10 November, 2015 | Recreating the Past

made with no human face as input

 

#FacesInThings

from an Instagram search #FacesInThings

This was going to be my first time using Kyle McDonald's face-detection library, I needed to calibrate so I gathered about 200 of the most recent #selfie tagged images from Instagram. The algorithm was able to detect faces in about 1 in 3. These were layered and alpha channels were blended evenly.

a blending of 60 #selfie images from Instagram

The faces didn't arrive pre-aligned. The first technical challenge was to align, scale, and skew each of the faces so they appeared on top of each other. The face detection algorithm provides eye and mouth locations, the following code aligns each of the images using those three points.

 

Ready to try this on #FacesInThings images. I got as many images as I could, this was about 2,500. In this case the algorithm identified faces in about 1 in 20 images.

 

 

#FacesInThings makes no distinction between human or animal, yet this image looks human. Also when I look for a gender, male and female both seem to fade in and out, very androgynous.

Brian pointed out that this image is almost more revealing of the algorithm itself instead of the images, so that led me to trying the algorithm on pure noise:

over 7,000 images like this were generated for source material

After hours of computation and 7,000+ images, the face detection identified faces in 47 images and produced this:

ghost in the machine

This is a visualization of the face detection algorithm; or a computer peering out from the void, learning a human face.

You can make out 2 dark marks for eyes (though very irregular), one for the mouth, and a brighter vertical strip as the nose, but totally absent are cheeks forehead or any face edge definition.

This one is especially eerie. It took so much computer power for just one barely visible face. Both #selfie and #FacesInThings blended images will arrive at an average of a human face after about 15 images. I'm not sure what it will take to get to that point with noise as an input, or if it's possible.

 

Thank you Zach Lieberman, Brian Solon, Daniel Shiffman