Testing Recognition

Eigen Faces 1 - 25

Eigen Faces 1 - 25

Analysis:
- As the number of Eigen faces increases, the more accurate the program can recognize faces.

- As the number of Eigen faces increases, the number of faces that are correctly identified increases but at a decreasing rate.

- From this graph, it appears that adding more Eigen faces would not improve the number of faces correctly identified

-I doubt adding 25 additional Eigen faces would improve the number of correctly identified faces. It appears that the benefit of adding more Eigen faces after 9, only provides small improvement.

- The more Eigen faces one uses, the more costly it becomes to compare faces, one gets improved precision. The less Eigen faces used yields less precision but can be compared to other faces with less cost of processing and storage space of each face.

-Eigen Faces yields poor results, only identifying 17 / 24 when most accurate. That is an accuracy of 70% Perhaps this could be improved by using color information or identifying other parts such as eyes or ears.

The Face Almost Matches ….

When running with 25 Eigen faces Image #2, shown on the left was identified as matching image #20 with an MSE of 638. Strangely the correct match was ranked 8th on the list of matches at nearly twice the MSE at 1151. This is not very accurate.

In this case with face 6 (on the left ) It was matched to face 7 (on the right). This was done using 25 Eigen faces. The correct match was ranked 2nd. And the MSE scores between 6 and 7 were and separate by a 10% margin. This is a more acceptable result. I also feel that these faces are somewhat similar.

Cropping and Finding Faces

 

Elf.tga

Min_scale = .45
max scale = .55
step = .02

This result has a false positive of the wall. I estimate that this is because the wall is very smooth. I could correct this by making some changes to the way I calculate my MSE. I could take color into account or check to see if the variance is close to the mean face. I experimented with the variance to eliminate false positives, but did not get any better results so my original implementaion is shown here. Please note that the face crop is the second best match after the wall.

Class0003.tga

Ran as expected no problems!

Min scale .45
max scale .55
step = .02

Class0004.tga

Class0005.tga

Class0006.tga

Facebook Pictures

Testing Recognition

Crop Result

Crop Result

Ran as expected no problems!

Min scale .45
max scale .55
step = .02

Ran as expected no problems!

Min scale .45
max scale .55
step = .02

The third best face is actually marks hand in his pocket.  The fourth best being the Washington sweatshirt kid.

Min scale .45
max scale .55
step = .02

The third best face is actually marks hand in his pocket.  The fourth best being the Washington sweatshirt kid. This took about 30 minutes to complete
Min scale 1.0
max scale 1.5
step = 0.1

I hypothesis that there are many errors do to my MSE calculation. I would like to enhance it some more.

Additionally, the people other then me were not in the training data and my get better results if I trained with their faces

\

I tred two MSE’s 400 and 1000, simply by using guess check. I expected to get similar results to what was found in recognition which was 70%

MSE 400:
My testing procedure was to validate every face smiling 1 - 23 against the non smiling userbase. It had 9 correct and  14 incorrect. Then I ran validate face 1 against each face and go no false positives. Face 1 was not mis-identifies as anybody else's face.

MSE 1000:
I used the same testing procedures, 21 correct identifications, and 2 incorrect. Although with I ran face 1 to test for false positives I got 16 false positives , meaning 1 was identifies as another person.

I would recommend using a MSE of 400 for best results