Phil McCarthy's Pareidoloop overlays randomly generated polygons on top of one another until facial recognition software recognizes a human face. Can't sleep, at SIGGRAPH! [via @Brandonn]


  1. Will version 2 have a feature that identifies instances of jesus and attempts to ebay them to goldenpalace?

    1. Version 2 will upgrade the facial recognition system from finding humans to finding cute kittehs.  Which will be promptly uploaded to one of the various kitteh repositories.  Such an influx of cuteness will decrease productivity and bring the internet to a crawl.  Soon after Skynet will go live.

      1. Given the nightmare fuel that this mechanism generates, I’m not sure that an overwhelming influx of the uncanny valley’s most unadoptable cat larvae will paralyze the world of the humans through cuteness

  2. It looks like the software detects something that looks like a face and then keeps the face-like aspects and then randomly generates more lines until it looks even more like a face than it did before, and so repeat. 
    So really it’s not surprising that at the end of the day it’s going to look like a face.  In Richard Dawkins’ book *The Blind Watchmaker*, many years ago, he had a computer program turn randomly branching trees into things that looked like specific animals. So I’m not sure what, if anything, this one shows.

    The description says: “What happens if you write software that generates random polygons and the software then feeds the results through facial recognition software, looping thousands of times until the generated image more and more resembles a face? Pareidoloop. Above, my results from running it for a few hours. Spooky.”

     It’s not clear to me what the ‘looping’ consists of.

    1. Yeah, that sounds to me like a genetic algorithm on images, where the fitness score is the facial recognition algorithm’s confidence value.

  3. I agree with above — any algorithm that builds on the weighting, or score, from the tool it’s claiming to “fool” is just demonstrating standard relaxation or hill-climbing techniques.  From a math and software point of view,  *nothing to see here, folks.*  

    1. Is it trying to fool it?  I thought it was supposed to randomly generate poly’s until the software recognised a face.  Seems to do what it says on the tin. *shrug*

  4. Instead of using face recognition software, it would be even more interesting to channel the multiple generation feedback loop in real time through the face-recognition areas of an individual human’s brain. Depending on how you set things up, it could be a hallucinatory experience, leading either to kittens, monsters, familiar faces, or maybe Alfred E. Neuman.

  5. Holy crap!  It’s Jesus!  Jesus lives in the computer!
    *looks at the next*
    Holy crap!  It’s Jesus!

  6. Seems like this process is coaxing the facial recognition to regurgitate the data it was trained with, meaning these faces are the program’s memories of actual specific photographs, or possibly chimera of several different people. I can’t decide if that makes it less, or more interesting.

  7. Some of them remind me of Roger Bacon portraits. They just need slabs of beef in the background.

Comments are closed.