These are 10 30x30 eigenfaces computed by my program
the left most one is the average face
This is the data plot of number of eigenfaces used against how many of a set of 24
images correctly recognized their smiling counterpart
Clearly as the number of eigenfaces increases, the accuracy of recognition
will also increase, since the original vector space will be more accurately represented.
However, it is also clear that the recognition to increasing number of vectors ratio
decreases as the number of vectors increases. Again, this makes sense given that the variance
of each subsequent vector will decrease, and proportionately less of the vector space will be
represented by each subsequent eigenvector. Therefore, striking a balance of about 9-15
eigenvectors seems to produce the most benefit without a huge number of eigenvectors.
Below are some examples of images which never correctly identified their non smiling counterpart
smiling 2, nonsmiling 20
smiling 6, non smiling 7
smiling 18, non smiling 20
Verify Data MSE = 1000
MSE: 562.162
Image 'smiling_cropped\01.tga' is a picture of nonsmiling_cropped\01's face.
MSE: 3017.66
Image 'smiling_cropped\02.tga' is not a picture of nonsmiling_cropped\02's face.
MSE: 698.834
Image 'smiling_cropped\03.tga' is a picture of nonsmiling_cropped\03's face.
MSE: 1502.96
Image 'smiling_cropped\04.tga' is not a picture of nonsmiling_cropped\04's face.
MSE: 1263.88
Image 'smiling_cropped\05.tga' is not a picture of nonsmiling_cropped\05's face.
MSE: 1673.84
Image 'smiling_cropped\06.tga' is not a picture of nonsmiling_cropped\06's face.
MSE: 2837.03
Image 'smiling_cropped\07.tga' is not a picture of nonsmiling_cropped\07's face.
MSE: 1754.3
Image 'smiling_cropped\08.tga' is not a picture of nonsmiling_cropped\08's face.
MSE: 509.438
Image 'smiling_cropped\10.tga' is a picture of nonsmiling_cropped\10's face.
MSE: 986.842
Image 'smiling_cropped\11.tga' is a picture of nonsmiling_cropped\11's face.
MSE: 880.779
Image 'smiling_cropped\12.tga' is a picture of nonsmiling_cropped\12's face.
MSE: 884.164
Image 'smiling_cropped\13.tga' is a picture of nonsmiling_cropped\13's face.
MSE: 1450.2
Image 'smiling_cropped\14.tga' is not a picture of nonsmiling_cropped\14's face.
MSE: 438.782
Image 'smiling_cropped\15.tga' is a picture of nonsmiling_cropped\15's face.
MSE: 914.903
Image 'smiling_cropped\16.tga' is a picture of nonsmiling_cropped\16's face.
MSE: 1008.84
Image 'smiling_cropped\17.tga' is not a picture of nonsmiling_cropped\17's face.
MSE: 2383.91
Image 'smiling_cropped\18.tga' is not a picture of nonsmiling_cropped\18's face.
MSE: 411.377
Image 'smiling_cropped\19.tga' is a picture of nonsmiling_cropped\19's face.
MSE: 1232.36
Image 'smiling_cropped\20.tga' is not a picture of nonsmiling_cropped\20's face.
MSE: 337.682
Image 'smiling_cropped\21.tga' is a picture of nonsmiling_cropped\21's face.
MSE: 1092.42
Image 'smiling_cropped\22.tga' is not a picture of nonsmiling_cropped\22's face.
MSE: 990.01
Image 'smiling_cropped\23.tga' is a picture of nonsmiling_cropped\23's face.
CORRECT: 11
false positive: 0 (none say correct for the wrong image)
false negative: 13
increasing mse to 1500 increases the number of correct elements by 6 without adding any false positives
increasing to 2000 adds a further 3 matches without adding any false positives
With off by one (verifying img i.tga with imgi+1.tga, nonsmiling) only produced one false positive,
with an mse of 1000
MSE: 730.067
Image 'smiling_cropped\10.tga' is a picture of nonsmiling_cropped\11's face.
Using input parameters min_scale,max_scale, step parameters of .45, .55, .01
adjacent is the version with the faces highlighted
To the right are some fine examples of facial recognition in images with 3 faces in them
Using input parameters min_scale,max_scale, step parameters of .45, .55, .01
Attempt to locate faces in class picture. Mixed results
computation time limited my ability to experiment with different scales
1.0 1.1 0.01 mark 20
This is my attempt to find myself in a random digital image.
Unfortunately, the illumination is not very consistant with the database images
and the algorithm is rather sensitive to this fact
I had similarly disappointing results with this image,
likely for the same reasons
0.65 0.75 0.01 mark 10
