How many images for good face recognition

asked 2016-05-17 08:28:17 -0500

atv gravatar image

I can't find a good answer for this.

Currently, my own experiments tell me i need at least 500+ images (currently i'm even thinking 1000 images) to get reliable face recognition, with lots of variation in illumination/facial hair, pictures close to the screen, far way, head tilted to the left or right, up and down etc. I am taking the images from webcam.

Now i'm approaching 2 problems.

  1. I suspect that to much variation will generalize the image and not help at all anymore. There must be a limit of diminishing returns.
  2. It takes an awful long time to start and train the program (10 minutes?) with 3 people and about 500 images each. Total image size is about 20Mb, and a saved model is about 56000 Kbytes. Would saving and loading the model be quicker on startup?

I would have hoped i would be able to do with much less (i can but then it is not so reliable i would be able to use it in a commercial fashion anymore). I don't see much difference using either Eigen or Fisherfaces. But maybe i need to keep more components, currently only doing 10 for Eigen.

Thanks for your comments! atv

edit retag flag offensive close merge delete


Just fyi, currently my datasets are not entirely even, i will make them all an even amount of images and retest.

atv gravatar imageatv ( 2016-05-17 08:59:27 -0500 )edit

This is a classical pattern recognition dilemma, nowadays with deep learning techniques the mass is typically more important - the more the better. If you don't have so much data then other methods might work as well (or even better).

Guanta gravatar imageGuanta ( 2016-05-17 11:05:48 -0500 )edit