CSE 455 - Project 4:Eigenfaces - Winter 2006
Testing Recognition
Using 10 eigenfaces, and training on the nonsmiling cropped images, I produced the following vectors:
The Average face: 
The Eigenfaces: 
I then repeated this recognition test for 1 through 32 eigenfaces. The following graph summarizes the results. This is before we were told to cut back on the face count.
Successful Matches vs. Number of Eigenfaces.
The plot shows that there is an increasing relationship between number of eigenvectors and the number of faces recognized. However, at around 10 Eigenfaces, the number of correct recognitions seems to cap out, as further eigenvectors do not increase correlations to the already unmatched faces. Since we only have 27 source faces to work with and any Eigenfaces created past 27 do not help the situation.
My attempt to find faces failed in some cases, but the actually person was always within the first couple of names. These are some examples:
| 01 matched as 22 |  |
 |
| 09 matched as 20 |  |
 |
As you can see the smiling faces of 01 and 22 have very similar features, so I think it is easy to see why the program got confused. It probably didn't help to have nearly the exact same patch of screen showing in the bottom right hand corner of the image. Other than that it's rather easy to see that their eyebrows would match almost perfectly, as well as the fact that they both seem to have the same lighting on their noses. I figured the MSE values for this comparison would be fairly close but actually 22 is 15000 while 01 is second on the list at 31000. The feature match to the eigenfaces just match much better to 22 than to 01. I find the second comparison to be a little more awkward as 09 has glasses and 20 does not. However if you look closly their features, shape of eyes and nose, they do closly match one another. The MSE here was within 1000 of being the correct match.
For some reason my algorithm was fooled into thinking the picture in the background was closer to a face than the actually faces.
This is even taking into account for the low texture areas. I tried modifying the scale but I still couldn't manage to get the face to be
first. However in the marked image you can see that the face was withing the first 3 values. The MSE of the picture is around 500 while the
actual face is above 700.
ME!
A picture of me taken when I first entered the Computer Science Department
The cropped image of my face from running my program with min_scale = 0.10, max_scale = 0.30, step = 0.01.
I chose these based on the fact that this is a fairly zoomed in portrait and my eigenfaces were only 25 pixels in size.
ME!
A picture of me taken when I first entered the Computer Science Department
The cropped image of my face from running my program with min_scale = 0.10, max_scale = 0.30, step = 0.01.
I chose these based on the fact that this is a fairly zoomed in portrait and my eigenfaces were only 25 pixels in size.
The following are several group face finding exercises.
The marking test seemed to work on most of these images. The scale for this was min 0.45, max 0.55 with a step of 0.01.
This marking also went very well and the scale was the same as the one in the above image.
This is the marking of the class group photo, I think overall it went very well as it does find most of the faces. There are a couple of mistakes. Many people in the back row aren't properly identified
this could be because the faces just aren't detailed enough to get caught by the algorithm. Some of the mistakes have me a little bit baffled, as I'm not sure why the upper right cabinet is selected, nor
the various selections on peoples clothing. I think if I had a little bit more time I would have tried to make this work on color hues because I think that would have solved most of the problems in this picture and
probably the elf picture as well. The scale I used for this photo was Min = 0.85, Max = 0.95, with a step of 0.01. I would like to find a more accurate scale but it take quite a long time to run when looking for
this many faces.
This marking also went very well and the scale was the same as the one in the above image.
This is the marking of the class group photo, I think overall it went very well as it does find most of the faces. There are a couple of mistakes. Many people in the back row aren't properly identified
this could be because the faces just aren't detailed enough to get caught by the algorithm. Some of the mistakes have me a little bit baffled, as I'm not sure why the upper right cabinet is selected, nor
the various selections on peoples clothing. I think if I had a little bit more time I would have tried to make this work on color hues because I think that would have solved most of the problems in this picture and
probably the elf picture as well. The scale I used for this photo was Min = 0.85, Max = 0.95, with a step of 0.01. I would like to find a more accurate scale but it take quite a long time to run when looking for
this many faces.
EXTRA CREDIT
implemented --verifyface