Computer Vision: A Modern Approach by Forsyth and Ponce

Check the manual of open CV library, there is a good explanation and C++ source code for the whole story.

Hi ,

Thanks for the input ! My understanding of PCA has improved tremendously but I have one last question. I understand that In the final stage of face recognition , the incoming image is projected onto the Eigenspace and the nearest distance to each face in the database ( transformed by the PCA matrix ) is computed. The minimum distance would then be classified as the matched face.

However, why can't the incoming image be projected onto the original face space and the nearest neighbor be the result of the matched face i.e we do not apply PCA at all. Is the purpose of PCA mainly to reduce computational load ?

I hope my question is clear. Thank you all.

Regards,

Haziq

You can do that as well (raw pixel classification). However, most of the information is redundant and can't be used reliably. When you use PCA  i.e when you project/multiply the input sample to Eigen space/Eigen vectors,  the most discriminatory information is capture (not just dimensionality reduction).. With this, you use most discriminatory as well as reduced feature set. 

Essentially: after PCA, your dimension of feature is say 'p' which is

The original paper by Turk and Pentland is probably the best on this topic.

Hi all ,

Thanks for the input ! I believe my understanding of face recognition using PCA is 99.99% complete.

Regards ,

Haziq