Project 4: Eigenfaces

Heather Langfeldt
CSE455 Computer Vision
Experiments

Recognizing Faces
Average Face:
EigenFaces:
Recognition Plot:
Discussion:	The more eigenfaces that are used, the better the recognition of faces. The cost of adding more eigenfaces is space and time. It would take longer to calculate and store the additional eigenfaces, and it would take longer for the algorithm to go through all of the eigenfaces each time to determine which face is the best match. According to my data, the most efficient number of eigenfaces is 5. Any more than 5 eigenfaces gives approximately the same percent correct recognition, but takes longer to compute. The most commonly mistaken faces are pictured below. I think many of these were inaccurately recognized by my code because each person changed one side of his face which unevenly distributed the features highlighted by the eigenfaces. Other mistaken identities could have occurred since faces were sometimes tilted in and out of the image plane. All but two of these mistaken faces were listed within the top five results.
Common Errors:

Cropping and Finding Faces
Crop 1:
Crop 2:
Find Faces:
Scales:	min_scale = 0.90; max_scale = 0.94; step = 0.2	min_scale = 0.74; max_scale = 0.80; step = 0.2
Find More Faces:
Scales:	min_scale = 0.74; max_scale = 0.80; step = 0.2	min_scale = 0.70; max_scale = 0.75; step = 0.005
Discussion:	Many of the false positives happened in regions of low texture. I tried to remedy this by eliminating the "faces" of which the center pixel was not a "skin" color. I used the true positives to determine the range of skin colors, but the next false positives usually came from textured object with a skin hue, such as tan shirts and hands. I was hoping for not too many false negatives because even if the most of the skin on the face was not within the range I had selected, it seemed probable that there was one pixel on the nose that was within the range and could center the square selection. I assumed this because light usually hits a person's nose, so there should be a few pixels with varying skin colors. Since finding the faces in the picture with me did not work well, I used a random neutral group picture from the class to crop. I have a feeling my own picture did not work as well because of hair color and the color of the buildings, both of which have a lot of texture and shadow to fool the algorithm.

Verifying Faces
Discussion:	I initially tried using a mse threshold of 6000, but many faces were appearing to be related, so I first inspected the mse values between true positive pictures, then took a look at the true negative values. Unfortunately, there is quite a large range where false positives and true negatives mix. The threshold that would recognize all of the faces would have to be greater than 1899, because this is the highest MSE value between the smiling and neutral picture of the same person, which means a mse threshold of 1899 would give no false negatives. On the other hand, such a high threshold would give the most false postives. A better MSE would be about 1000, which excludes four of the highest MSE scores, but reduces the number of false positives. As far as I could gather, with a threshold of 1000, the false negative rate is 4/33 and the false positive rate is about 102/1056.