Is there any way I can make the FaceRecognizer's training faster?

asked 2017-09-12 05:13:02 -0500

Gensoukyou1337 gravatar image

I'm developing for the Android platform and, as of my current understanding, I'd like to use a YAML generated by training a FisherFaceRecognizer in my computer (via Python) to be loaded into the FisherFaceRecognizer in the Android app I'm testing it in.

In my last time training the FisherFaceRecognizer in my computer with 1923 images resized to 120x120 from the Wiki part of the IMDB-WIKI dataset divided into eight labels by both gender and age group, it took 1506 seconds in total. I feel as if that would be too slow, especially if I need to train the FisherFaceRecognizer with many more pictures.

My question: Can I make the training process any faster other than just buying a computer with a faster processor?

edit retag flag offensive close merge delete


it took 1506 seconds in total.

no, that is NOT slow, more average.. also, for something as "static" as age & gender recognition, you're supposed to do the whole training only once.

you can try with smaller images, like 90x90, but hmmm.

berak gravatar imageberak ( 2017-09-12 05:53:02 -0500 )edit

Then I guess it's just me. Also, let's say I have eight categories, based on two genders and four age groups. How many face images do I need per group, at minimum, for a robust recognition? Because right now I have around 250 images per category.

Gensoukyou1337 gravatar imageGensoukyou1337 ( 2017-09-12 09:55:03 -0500 )edit

hmm, maybe you should train 2 seperate models, one for gender, the other for age

berak gravatar imageberak ( 2017-09-13 02:30:06 -0500 )edit

I thought I had to train 3 separate models, one for gender and the other two for age separated by gender.

Gensoukyou1337 gravatar imageGensoukyou1337 ( 2017-09-13 04:54:32 -0500 )edit