Facial Recognition with Linear Support Vector Machine

606 images, along with 606 labels (the names of the subdirectories the images were contained in) were used to train a linear SVM. The SVM was then tested on 1818 images that do not include the training images.

The linear SVM matched the 1818 test data to the correct labels, better than 80% of the time.

With more training data, the accuracy goes up to 92%. I am quite impressed! I didn't really think faces were linearly separable and just toyed with it to see what I could get.

I've been trying the polynomial and RBF kernel SVM with less success. There are parameters to tweak for them, which if properly done *should* give me better results than linear SVM. I just haven't figured out how.