The 10 eigenfaces:
Average face:
The 10 eigenfaces:
Face recogniztion: Plot of the number of faces correctly recognized versus the number of eigenfaces used
1. Basically, the number of faces correctly recognized increases with the number of eigenfaces. However,
it is not monotonically increasing. And at some large number of faces, the curve
becomes "saturated". There is no clear answer of how many eigenfaces should one
use, but from the plot I think 21 is a good trade-off between accuracy and
speed.
2. Recognition errors example:
This is one pair that was mutually mis-recognized, i.e., Song is recognized as Hazen, and Hazen is recognized as Song as well. Actually these two images are quite alike, that's why the mistake is reasonable. When using 10 eigenfaces, the correct answer of Hazen appears highly in the sorted results (#5), but the correct answer of Song does not appear highly in the sorted results (#12).
Parameters: min_scale = 0.45, max_scale = 0.55, step = 0.01
Final results: scale = 0.47, x = 138, y = 126, mse = 233.89
Original image: 
Cropped image: 
Parameters: min_scale = 0.55, max_scale = 0.65, step = 0.01
Final results: scale = 0.55, x = 76, y = 47, mse = 513.48
Parameters: min_scale = 0.95, max_scale = 1.05, step = 0.01
Final results:
Face 1: scale = 0.95, x = 72, y = 94, mse = 102.41
Face 2: scale = 1.05, x = 312, y = 87, mse = 113.91
Face 3: scale = 0.97, x = 187, y = 92, mse = 191.21
1. The scaling factors are determined by inspection. Specifically, they are:
Cropping the Aseem.tga:
min_scale = 0.45, max_scale = 0.55, step = 0.01
Cropping other images:
Marking the Group1.tga:
min_scale = 0.95, max_scale = 1.05, step = 0.01
Marking other images:
2. For Aseem.tga, if the program is asked to find 2 faces, then it gaves this result, which causes false positive:
For Group1.tga, if the program is only required to find 2 faces, then the result is false negative, since one face has to be omitted:
In the former case, the area with texture the most similar to a face is selected as the second face; In the latter case, the two faces with the strongest similarity to the eigenfaces used are selected, while the weaker one is omitted.