Tom Anderl

Project #4
Vision Winter 2003

Answers to the questions:

Recognition with cropped class images

1. 19 out of 22 faces were correctly recognized
2. For the 3 faces that were not correctly recognized, their positions were 1, 6, and 3 (where 0 is the index of the face the program thinks it is
3. There does not appear to be a good correllation between the sex of the face and the sex of the top matches. Every match had at least one male and at least one female within the top 4 matches.
4. It depends on whether I'm going to be searching over the entire user base. I would like to limit the faces used to construct the face space to the faces that I expect to be looking for. Judging from the results of the image recognition using a face space of the entire class, I would say that a face space biased towards the users it's trying to recognize works better.
5. I actually think the opposite, since constructing the face space from the user set would yield a face space sensitive to the differences found in the users. A more general face space would have to be sensitive to other factors, and thus would probably not be as good at picking out the differences between the users.

Recognizing undergrad faces

1. My program consistently recognized esp, melissa, and tamoore, and no one else.
2. I think that the incorrect results are reasonable, for example, it confused these two images:
   
3. This performed much worse than the previous recognition experiment. Possible reasons for this poor performance are:
   • More generic average face. The average face used here is more different from the images it's trying to recognize, so the mse of an image to the average face is probably higher.
   • More generic eigenvectors. The eigenvectors used here are more general, and are thus, less sensitive to the differences seen in these specific images.
   • Different user set from training set. The user set was not exactly the same as the training set, so the trained faces were not as keen to the differences in the user set.
4. The number of eigenvectors did not seem to affect the results at all. Note the fact that there were always exactly 3 correct matches.
5. It's best to use the least number of eigenfaces because that means less computation and less storage space.

Cropping the undergraduate faces

1. I ended up using min_scale = 0.2333, max_scale = 0.3, step = 0.03333.
2. About half of the images showed up correctly. Here is a sample of the results:

   ratliffn	adeakin	afuchs	aday

3. You end up with too big of a cropped image if you use a min_scale that's too small.

Faces in a group photo

1. Results of the face finding can be found here.
2. I used the following parameters for finding faces in the pictures:

   	min_scale	max_scale	step	mse_thresh	texture_thresh
   Beach Boys	0.9	1.1	0.025	2000	450
   Sieg Dwellers	0.75	0.9	0.05	5000	100

   *Note: The mse_thresh and the texture_thresh are described on the face finding page.
3. The program does find a number of false positives for faces and often will not find a face where there is one. In some cases, I think this is because the face in the picture is quite different from the faces from the training set. For example, in the Beach Boys picture, the member with the beard and glasses does not get found. This is likely because the beard and glasses give him a different enough look that he is unrecognizable by the program.