Revision history [back]

I am not familiar with the fisherfaces method, but it is pretty obvious that using a more diverse training set would help the performance of your application. Each race has its unique features so you'd gain by teaching them to your classifier.

There are dozens of face datasets on the web, but finding one that is labelled with age and gender might not be easy. You might end up having to label a dataset yourself. Start by browsing this website, it has a collection of different face datasets.

Good luck.