Genetic Algorithms: Evolving a human face

I work for a face recognition company (http://www.pittpatt.com) so one friday afternoon I got the crazy idea of creating a genetic algorithm to evolve human faces using our detector as the fitness function.

See: http://lbrandy.com/blog/2009/04/genetic-algorithms-evolving-human-faces/ for the full write-up

This is those results.

Category:

Science & Technology

Tags:

• face
• detect
• recognition
• pittpatt
• lbrandy
• genetic
• algorithm

[ernesto50]

congratulations, it seems to get stuck in a local minimum

[ludamad]

It's evolving against their face recognizer, I guess it has a measure of likeness to a face used as the fitness function.

[Mjhond]

I am very happy to see the vidoe after you give this o one friday afternoon I got the crazy idea of creating a genetic algorithm to evolve human faces using our detector as the fitness function.

[Ondelendo]

I Love The Video It Can Increase My Knowledge so one friday afternoon I got the crazy idea of creating a genetic algorithm to evolve human faces using our detector as the fitness function.

[anakmudajaman]

Steady I Really Like This Video so one friday afternoon I got the crazy idea of creating a genetic algorithm to evolve human faces using our detector as the fitness function.

[bebeheuy]

Good, I like that you share this video, I wish success always so one friday afternoon I got the crazy idea of creating a genetic algorithm to evolve human faces using our detector as the fitness function.

[imegatrone]

Nice Video That You Share , So Very Nice Thanks You so one friday afternoon I got the crazy idea of creating a genetic algorithm to evolve human faces using our detector as the fitness function.

[willamricard]

I Really Like The Video From Your so one friday afternoon I got the crazy idea of creating a genetic algorithm to evolve human faces using our detector as the fitness function.

[bundawartini]

Your Video Is Very Useful Sharing so one friday afternoon I got the crazy idea of creating a genetic algorithm to evolve human faces using our detector as the fitness function.

[TiagoTiagoT]

And about the local maximum/minimum; perhaps you could make it so the frequency and/or intensity of mutations and crossovers is inversely proportional to the difference between the best scores of this generation and the previous' ?

Perhaps with a delay or moving/weighted average instead of just using the current difference from the last, to reduce feedback loops leading into extreme oscillations, and smooth out the changes in intensity

[TiagoTiagoT]

Btw, for doing the crossover; wouldn't it be just a matter of picking two of the best individuals of each generation, and then doing it like you do with plain mutation but instead of being random it's the average of the color for the pixel in the two parents (probably you wanna have each mating produce at least two new individuals, one is parent A with some pixels mated with parent B's and the other is parent B with the same pixels mated with parent A's) ?

[TiagoTiagoT]

An interesting approach to see what the detector thinks a face should look like.