Eigenface #

jason.dang · jaydang@cs#

note: if for any reason you don't want your picture to be on here, please let me know

 

$Bingo, found it! $

recognizing cse455 students

say cheese?

what's gotten in2 u?

Here are the seven victims caused by the 22 non-smiling cse 455 students:

Here are three "most-wanted" that eigenface.exe couldn't seem to find:

Here are the top five that are mostly like to be on the "list":

Here is record of the top 10

 

 

Recogn.

1.

2.

3.

4.

adeakin

 

P

M

F

F

M

alissah

 

P

M

F

M

M

amiratuw

 

P

M

F

M

F

crosetti

2

O

M

M

F

F

ddewey

 

P

M

F

F

M

djj

 

P

M

M

F

F

esp

 

P

F

F

M

M

galen

 

P

M

F

M

F

jaydang

 

P

M

F

F

M

jwkim

 

P

F

F

M

M

margaux

 

P

M

F

M

M

mbixby

 

P

M

M

F

F

melissa

 

P

F

M

M

M

merlin

 

P

M

F

F

M

mhl

6

O

F

F

M

M

paullarp

3

O

F

M

F

M

 

Using my eigenface program, I was able to recognize 19 out of the 22 faces.  The 3 that were not recognized came pretty closed ... average out to be around 4.

There isn't a pattern relating the males and the females.  However, one interesting observation that I made was that there around 5 people that keep on making it to the top five list.  

Even if there are thousands of users, I still wouldn't use the entire set to compute the face space.  The reason for this is because once you have a pretty good variety of sample, that sample can be used to the rest of the people.  Also, running more picture take longer execution time.

When computing the eigenfaces, it's better to use a face set independent of the user because it's a more effective way to test the program.  Also, in real-life, people's looks are varied from image to image. 

 

 

 

$Where did everyone go? $

recognizing cse undergrads

 

12 undergrad eigenfaces

    average:

5 undergrad eigenfaces

      average:

10 undergrad eigenfaces

    average:

15 undergrad eigenfaces

    average:

20 undergrad eigenfaces

average:

 

Because of the angle and shape of their face, these five faces seem to be the ones that appears on the top five list.

Faces that were hard for the program to recognized due to facial feature and orientation.

 

Unlike the previous test, this time only one face come close to a match and the others are not even near at the top ten.  
The three main reasons that the program performs so poorly is that:
  1. The five face shown above are very closed to the average and therefore seem to always match.
  2. The eigenfaces are extract form a different set of data
  3. When smiling, the shape of the face are shifted and therefore harder to for the program to recognize.                 
As the number of eigenfaces decreased, the MSE decrease...hence, recognizing that the face is a closer match.  However, the list seem to stay quite the same.

 

note: if for any reason you don't want your picture to be on here, please let me know

 

CSE 455 Computer Vision Winter 2003