| Testing Recognition with Cropped Class Images |
| The eigen faces generated from the set of cropped nonsmiling CSE 455 students |
 |
 |
 |
 |
 |
 |
 |
 |
| 0 |
1 |
2 |
3 |
4 |
5 |
6 |
Average Face |
|
| click [ HERE ] to see how the program recognized smiling students |
| (1) | The program recognized 19 faces correctly. |
| (2) | For the incorrect predictions, the average position of the correct face in the recognition list was 4.3ish |
| (3) | The program matches female and male faces independant of any order. The program attempts to find faces in the
space spanned by the eigen faces. Though members of the same gender may have similar values in the face space, facial recognition is
independent of whether the previous face was found proximal to the current face in this face space. The recognition data does not showany an order/gender pattern -- it appears completely random. |
| (4) | In the real life scenario, we should not use the entire set of users to compute the face space. It would be better to get a set of random faces that are fairly different from eachother so that they span a large amount of the face space. Using more faces would not help much - if at all. If we used more faces much of the information would be redundant since they would fall within the already determined space. Using more faces (after a certain point) will only have a minimal affect on the average face. In addition, computing the face space is computationally expensive - therefore, the fewer faces we chose to
define the space, the better. |
| (5) | We want to compute eigen faces from a face set that is independant of the user set. We hope the space spanned by the eigen faces is a good representation of the actual face space. Then, when projecting two faces into this space, we determine they are the same if they are near eachother. When the user set is used to define the face space, we are projecting the user faces into a space that the user helped define.
*** Write more here **** |
| |
| Recognizing the Undergraduate Faces |
| The eigen faces generated from the set of cropped undergraduate students |
 |
 |
 |
 |
 |
 |
| 0 |
1 |
2 |
3 |
4 |
5 |
 |
 |
 |
 |
 |
 |
| 6 |
7 |
8 |
9 |
10 |
11 |
|
| | Average Face |
| [ 5 eigen faces] click [ HERE ] for output using 5 eigen faces. |
 |
| [10 eigen faces] click [ HERE ] for output using 10 eigen faces |
 |
| [12 eigen faces] click [ HERE ] for output using 12 eigen faces |
 |
| [15 eigen faces] click [ HERE ] for output using 15 eigen faces |
 |
| [20 eigen faces] click [ HERE ] for output using 20 eigen faces |
 |
|
| (1) |
There are 4 faces matched correctly and 14 faces mismatched. |
| (2) |
These incorrect identifications are NOT reasonable. Everyone looks like melissa garcia. Even Linn! |
| |
|
| (3) | The pictures of all the undergraduate students were taken under varied conditions. Some are smiling, some are in
shadow, some faces are slightly off center, some have glasses. All these are temporal things that change how the face appear to the computer. Our
program can't distingush between these temporal features and more permanent features, and therefore draws incorrect associations between people
that have very different appearances. The previous method - using the class pics - gave better results. In this case, we used the same camera
for every picture, the lighting was the same, etc. In addition, we know that the face space is well defined where we projected the class faces (in
the previous method) because we used the class faces to define the face space. |
| (4) | Increasing the number of eigen faces SHOULD make the results better - faces are defined by more unique equations and
there is less chances that two unequal faces have the same coefficients in the space. HOWEVER - I think because of a lot of inconsistancies in the pics
we were using - I actually got 0 matches when I increased the number of eigen faces to 20. |
| Cropping the Undergraduate Faces |
| While experimenting with min, max values, I got a couple "faces" that were too large:" |
 |  |  |
|
| But when I set min_scale=1.0 and max_scale to 3.5, I was able to get a face every time. I am also right arount 10 seconds,
when the step is set to 0.1. I am consistantly around 10 seconds when I scale from 1.0 to 3.0 with a step of 0.1. |
|
|
| click [ HERE ] for details about the results and the run time. |
| Finding Faces in a Group Photo |
|
|
For the Royal Tenebaums image I set min_scale = .3, max_scale = .7, and step = .1
For the picture with my friends, I set min_scale = .2, max_scale = .4 and step = .1 (I also drew more boxes than I needed)
I had a hard time finding pictures of my friends / family -- this was the best I could do (at least that was digital). It was difficult because not
everyone was looking straight ahead. |
