EIGENFACES

Extra Credit

-Implemented 3rd whistle -- speedup in Faces::eigenFaces()

-Implemented EigFaces::verifyFace()

_______________________________________________________________________________ 

Testing recognition with cropped class images

average face                eigenvectors 1 to 10

1.

Generally, as the number of eigenfaces increases, the number of faces recognized 

increases. The increase in the number of faces recognised is steep in the initial portion

of the graph and levels off towards the end.

The more eigenfaces one uses, the longer one takes to generate them. One could plot

the graph as shown above from 1 to the maximum number of eigenfaces (total number

of training images) and use it to find the number of eigenfaces that give the highest 

number of faces recognized. 

2.

The mustakes in the shell above are reasonable . In both cases, the correct

face is ranked second.

_______________________________________________________________________________ 

Cropping and finding Faces

Original Images: 

Cropped Faces:                                        

min_scale = 0.45, max_scale = 0.55, step = 0.01 are used as parameters for both images.

______________________________________________________________________________________

Original: 

Marked Faces:

min_scale = 0.5, max_scale = 1.1, step = 0.1 are used as parameters for both images.

In the right image, the false negatives may be due to the lack of faces with dark facial hair and

dark-rimmed glasses in the training samples. It is puzzling why the woman at the bottom right

is not correctly marked. The false positives fall in regions with low texture. It seems that multiplying

the MSE by the distance between a face and the mean face, and then dividing by the face's variance

does not fully solve the problem of low texture areas.