Ask Your Question
0

Fisherfaces improve classification?

asked 2013-05-13 05:09:02 -0600

forthtemple gravatar image

updated 2013-05-13 05:10:24 -0600

I'm using Fisherfaces for face classification of race and gender and I cannot get it better then about 80%. Wondering if anyone has any tips on improving it. Note I'm also trying various matlab fisherface scripts aswell and all are about 80% accurate.

Note that the things I've tried is:

  1. Increasing the name of training images. But I find 80 images is about the maximum where there is only marginal improvement when it is more than that.

  2. Instead of using the best picked face match, instead get the average of the top three. Eg if face one is male, face 2 is female and face 3 is female then choose female as the gender. But I find this gives worse results than just picking the best matched face.

  3. Changing normalisation, contrast etc of the images. But I find this only has a marginal effect on the accuracy.

edit retag flag offensive close merge delete

Comments

Hmm, have you considered using a different algorithm? Fisherfaces are of course a kind of baseline algorithm for face recognition but nowadays already a little bit out-dated.

Guanta gravatar imageGuanta ( 2013-05-14 03:33:15 -0600 )edit

Is there any algorithms you would recommend for classification?

forthtemple gravatar imageforthtemple ( 2013-05-15 16:17:01 -0600 )edit

Sry, I don't have any concrete suggestions, make a literature search on recent papers and see against which algorithms they compare nowadays.

Guanta gravatar imageGuanta ( 2013-05-15 16:26:46 -0600 )edit

2 answers

Sort by ยป oldest newest most voted
1

answered 2013-05-14 04:14:01 -0600

Like guanta suggested, fisherfaces and eigenfaces do their work, but lately more interesting algorithms have been released for matching known images to a new input image (basically what you do in face recognition).

Also 80 images is not much. Indeed you get marginal increases, but getting like 1000 images of each reference face, in all kind of circumstances, yields better results. You should vary your data, different lighting, different moods, different facial hair conditions, ... It is an exhaustive piece of work to get a robust system.

edit flag offensive delete link more

Comments

I tried 200 training images and got no improvement. Also I found that eigenfaces is no good for classification. I think eigenfaces is good for face recognition while fisherfaces is better for face classification.

forthtemple gravatar imageforthtemple ( 2013-05-15 11:05:35 -0600 )edit

classification is basically the same as recognition here, you just need to get much larger datasets first to seperate the data well. PCA and eigenfaces perform pretty well here. In order to capture enough variation between for example male and female, you would need about 5000 males and 5000 females to get a decent discrimination between both, Humans, and especially faces, have a very large intra class variability that you need to cope with!

StevenPuttemans gravatar imageStevenPuttemans ( 2013-05-15 15:33:51 -0600 )edit

Its a shame no one has a created a haarcascade_gender.xml classifier. Would be good to check if it is better than 80% that I'm getting with fisherfaces with just 80 training images.

forthtemple gravatar imageforthtemple ( 2013-05-15 16:14:34 -0600 )edit
0

answered 2015-08-23 17:11:48 -0600

jeremyrutman gravatar image

updated 2015-08-23 17:13:01 -0600

I have tried to reach the '98% accuracy' claimed in the fisherface demo but so far hardly break 80% for gender detection as you can see from the graph below of % correct classification vs. number of training examples for each class.

accuracy vs N training

This is using the LFW face database, registered/scaled with 'deep funneling' as provided at the LFW site. Anyone find anything better? Would larger training sets help?

edit flag offensive delete link more

Comments

The reported efficiency of the fisherfaces algorithm is on the person identification task using the complete AT&T database. There it does work perfectly well. For gender classification I think 80% is already a good starting point. I am not convinced that Fisherfaces will do the trick there to get 98%.

StevenPuttemans gravatar imageStevenPuttemans ( 2015-08-31 08:14:01 -0600 )edit

Question Tools

Stats

Asked: 2013-05-13 05:09:02 -0600

Seen: 1,719 times

Last updated: Aug 23 '15