CSE576 - Computer Vision

Michael Eckert

0. Average Face and the 10 Eigenfaces

The Average Face
Eigenfaces 0 - 4
Eigenfaces 5 -9

1. Recognition with cropped class images

My plot:

Around 20 Eigenfaces or so the number of recognization errors stabilizes -- using a higher number does not result in any gain.

Also around 8 Eigenfaces or so the curve's ascent is becoming strongly less. Using something about 10 Eigenfaces should be a good tradeoff.

A couple of recognization errors (with 21 eigenfaces):

These pairs with recognization errors actually show a relatively similar shape of eyes, nose and mouth.

2. Cropping and Finding Faces

My error-function is mse * dist^4 / var; I found this worked better than the suggested mse * dist / var (esp. in the pumpkin picture)

2.1 cropped Aseem

2.2 cropped me (parameters used: 0.60 - 0.65 in 0.01)

2.2 cropped Westerwelle (parameters used: 0.20 - 0.30 in 0.01)

2.3 cropped group1.tga (parameters used: 0.80-1.00 in 0.02 mark 3)

2.4 cropped Group picture (parameters used: 0.70-0.90 in 0.02 mark 5)

False negatives: both false negatives suffer heavily from jpg-compression artifacts. Additionally in my face my eyes are almost closed while the other unrecognized face is slightly rotated.

False positives: Both false positives have in the errors they produce a quite acceptable gap to the three faces. So they mainly result from the algorithm's inability to classify the above false negatives. Without the jpg-compression artifacts, the gap would probably be much bigger.