Jim George CSE 455 Assignment 4

Winter 2009

Recognition

Recognition Questions

1. Plot trends.
The plot shows a strong increase in value when adding from 1 to 10 EigenFaces. However after this there is no value gained what-so-ever. It is clear that using about 10 faces in this case is optimal
2. Recognition Errors
I have included below two recognition mismatches from the algorithm:

These two do look very similar, both in luminance town and face placement.


These however seem very different. perhaps it is the angle of the mouth that made them similar.

It seems possible to conclude that the "patterns" the algorithm is discovering are seemingly very different than our perception, so it is hard to judge for sure with these small examples. In general, however, the plot indicates that even with just 10 eigenfaces we are able to recognize smiling vs. unsmiling fairly successfully.

Cropping & Finding Faces

1. Elf image.
This annoying image would not stop catching the top of this man's head for the best match. If I added 3 faces I was able to capture him and his startled offspring.

Scale of .6 with the 10 EigenFace file.


2. Self.
in this fairly picture of myself I was able to locate my face as the third ranked region, with an scale of .4 on the 10 EigenFace file



3. Groups.
In both smiling and non smiling groups of three i was able to identify all three faces

Scale: range from .61 to .7 step .01


Scale: range from .51 to .61 step .01

In the large group photo I had some trouble locating results for everyone.

Scale: .8 through 2.0 step .05

Scale: .6 through 1 step .05


The false positives and negatives in these photos are most likely due to the weak nature of the EigenFaces algorithm against luminance differences. The algorithm performs very well when lightning conditions are unchanging, such as the examples between smiling and somber classmates. However even in the same classroom with the camera reversed towards the lights the performance degrades. Also textures such as the one in the family portrait seem to commonly catch the algorithm because with textural variance it is likely that a portion of it will end up statistically similar to a face.

Verify Faces

To verify a face I ran the verifyFace method on the nonsmiling user base against their smiling images. My "MSE" thresholds ranged from .001 to .0095 and I sampled both False Positives and False Negatives at a granularity of .0005
The best threshold I computed to be .0055. I determined this by examining the plot below and seeing their intersection. This is where both False negatives and False positives are at their lowest, which is only 8%.

NOTE: Because I experimented with dividing the overall image with the norm of the image my MSE values are quite are at a different scale than that of the standard MSE algorithm