Monday, August 6, 2012


This is a program that satisfies, at a very simple level, the creation/evaluation loop that I felt would be necessary for any machine to make something we would be able to find creative. It randomly places triangles  until it finds a face using face detection software. A few results are pictured above. I see a lot of character in these faces that I think wouldn't have shown up in a more direct synthesis from a face model.
The author is Phil McCarthy.
The face detection method is the really rich part of this program that does most of the heavy lifting. It is based on this paper: High-Performance Rotation Invariant Multiview Face Detection by Chang Huang, Haizhou Ai, Yuan Li, and Shihong Lao, from Tsinghua University. The paper advances both machine learning techniques and optimiztion methods to make it fast. The training set is Labeled Faces in the Wild. The faces in the training set form the algorithm's idea of what a face should be.
The program was inspired by Greg Borenstein running a face tracker on the flickr pool Hello Little Fella!, and by Evolution of the Mona Lisa by Roger Alsing.

The name of the program, Pareidoloop, comes from the word pareidolia used to describe our tendency to see faces in the clouds, or in the rocks.  I put this poster by Bev Doolittle on my bedroom wall as a kid, that plays with this idea:

(The poster came from Ranger Rick magazine, which I had a subscription to for years.  I learned a ton of science from that magazine; it built up a lot of the structural framework in my memory that later ideas I was taught about ecology, botany, and so forth attached to.)

No comments:

Post a Comment