Peter Beckfield
CSE 455
Project 4, Eigenfaces for Recognition


Testing recognition with cropped class images

Average Face

Eigenfaces


Accuracy vs. Eigenfaces used


1. An increasing number of eigenfaces seems to have diminishing returns (which makes sense, since the eigenvectors with lower eigenvalues would have less of an impact on the image as a whole).
It looks like using eigenfaces equal to about half the number of training images gives acceptable results while still keeping the computation quick.


Cropping and finding faces




1. I used .2 for min scale, .4 for max scale, and a step of .05 for the non-smiling group image. I was getting worse results for the
smiling images, so I changed to .2 min, .52 max, .04 step.
2. There were many false positives, many of which were due to the fact that I ran out of time to implement the removal of intersecting faces.
Because I didn't have this, I wanted to make sure if one person's face had a very small MSE, other faces would still get boxes.
Even disregarding the intersecting boxes, my results weren't stellar. Oftentimes only two of the faces in the group would be detected,
and then there would be a patch of boxes at some random location.