Project 4

Malcolm Bixby

Part I:

1. Calculated 19 faces correctly. 3 Incorrectly.

Face '..\..\allfaces\files\class_smiling_cropped\crosetti.jpg' recognized as being closest too:

0: ..\..\allfaces\files\class_nonsmiling_cropped\paullarp; MSE: 11474.9

1: ..\..\allfaces\files\class_nonsmiling_cropped\crosetti; MSE: 13384.7

and

Face '..\..\allfaces\files\class_smiling_cropped\mhl.jpg' recognized as being closest too:

0: ..\..\allfaces\files\class_nonsmiling_cropped\melissa; MSE: 18586

1: ..\..\allfaces\files\class_nonsmiling_cropped\margaux; MSE: 19253.1

2: ..\..\allfaces\files\class_nonsmiling_cropped\alissah; MSE: 23861

3: ..\..\allfaces\files\class_nonsmiling_cropped\amiratuw; MSE: 31874.5

4: ..\..\allfaces\files\class_nonsmiling_cropped\ddewey; MSE: 39507.8

5: ..\..\allfaces\files\class_nonsmiling_cropped\adeakin; MSE: 43503.8

6: ..\..\allfaces\files\class_nonsmiling_cropped\mhl; MSE: 45869.8

Face '..\..\allfaces\files\class_smiling_cropped\paullarp.jpg' recognized as being closest too:

0: ..\..\allfaces\files\class_nonsmiling_cropped\alissah; MSE: 8043.42

1: ..\..\allfaces\files\class_nonsmiling_cropped\ddewey; MSE: 18170.5

2: ..\..\allfaces\files\class_nonsmiling_cropped\margaux; MSE: 30119.1

3: ..\..\allfaces\files\class_nonsmiling_cropped\paullarp; MSE: 32758

Average: off by 4 for wrong ones.

3) Of successful matches:

10 leading males followed by other males

1 leading male followed by females

3 leading females followed by other females

5 leading females followed by males

Secondary:

5 secondary males followed by other males

3 secondary femles followed by other females

7 secondary males followed by females

2 secondary females followed by males

Males appeared to be followed more often by other males but there otherwise doesn’t appear to be a pattern.

4) No. There would be no need. Given that humans are by and large similar, a small sampling of the users would give an acceptable average face, additional faces would only be modifying the average face slightly and not contribute much to the overall results. There are only so many faces that can fill a spanning set, eventually a limit would be reached where any additional faces aren’t changing the average significatnly enough to affect the results.

5) By using a face set independent of the user set, there is likely to be variation among the faces used to calculate the eigenvalues which should yield a more accurate set.

II.

1) nosmile_eig5.face recognized 2 faces

2) nosmile_eig10.face: recognized 3 faces

3) nosmile_eig15.face: recognized 2 faces

4) nosmile_eig20.face: recognized 3 faces

5) Processing time is shorter with fewer eigenfaces.

III.

People with glasses (or shiny cheekbones) tended to match the closest.

Tamoore, Mhl, and Melissa were recognized the most.

The incorrect identifications are not reasonable, as Melissa was recognized as the closest match in most every case.

1. The average face results in a front-framed image of a person. It is likely that those pictures of people that are also framed more or less exactly in the middle of the image will match closest to the average, as is the case with the cropped image of Melissa.

2. The hues and shadows vary widely in the ugrads_cropped file while the hues and shadows are similar for all pictures taken of students in class.

3. People with glasses or whom are smiling may through off the results.

4. In my case, changing the number of eigenfaces appeared to have little effect on the results.

IV. Nope, never did get the crop to work. Mark either for that matter.

1. Used all sorts of min/max scaling. Most often it was a neck or shirt that was cropped.

2. The program seems to have a propensity for blue shirts.

1. I stepped incrementally through a range and would re-run the findface within the range that produced the lowest eigenvalues but with a smaller step size.

2. Errors? Heaps of 'em. Program seems to like walls, shirts, and adam's apples before faces...

On to the images:

A selection of eigenfaces: