Testing recognition:

Average face:

Eigenfaces:

Accuracy of recognition:

Questions:

1. There is not a clear best number of eigenfaces to use, although additional faces beyond about the first half do not help much.
2. Here are two of the recognition errors for 23 eigenfaces:

Test Picture

Algorithm Match

Correct match

In the first set, the matched picture does seem more similar in terms of facial feature shape, especially in terms of mouth shape. The second set less clear, the test picture does not seem terribly similar to the algorithm picture. In neither case was the correct answer ranked highly -- in the first set, the correct answer was #16, and in the second case the correct answer was #20.

Cropping and finding faces

Results

1

 ===>

2

 ===>

3

 ===>

4

 ===>

Questions:

1. Parameters:
  1. Aseem -- range 0.45-0.55, step 0.01.
  2. Perry Farrell -- range 0.50-0.55, step 0.01.
  3. Group1 -- range 0.89-0.95, step 0.01.
  4. Goodfellas -- range 0.45-0.55, step 0.01.

2. The only real false positive is in the Goodfellas picture (#4). As can be seen, the guy on the right (Joe Pesci) fools the algorithm because he is not straight-on to the camera. Other than that, the goodfellas picture was actually pretty easy due to the large black regions. One note -- with picture #2, I actually had to crop it just a little to take out some low-texture regions that fooled the algorithm (this despite the factors I added in to penalize distance from the average face and low variance).