Project 4 Artifact

Mark Agoncillo

CSE 455 – Computer Vision

Project 4 Artifact

Part 1 : Testing Recognition with Cropped Images

Average Face :

nonsmiling_cropped:

Questions:

1) As the number of faces used increased, the number of faces recognized also increased. Five faces seems to be a logical choice for how many to use, as the greatest increase in number of faces recognized occurs between 3 and 5 faces used.

2) These images were proved more difficult to recognize with the program. The mistakes were pretty reasonable, however. The correct answer appeared reasonably highly in the sorted results.

Part 2: Cropping and Finding Faces

Cropping the Elf picture

findface on elf.tga with min_scale = 0.45, max_scale = 0.55, step = 0.01

When I performed findface on elf.tga, the cropped image that resulted was of the man’s eye. Since the average face was of mostly young adults, it recognized the man’s face more than the baby’s face. It also appears that the scale was not small enough to recognize the entire face, so it cropped only the eye.

Cropping my Self-Portrait

findface on mark.tga (below left) with min_scale = 0.45, max_scale = 0.55, step = 0.01

Cropped face image:

Cropping my self-portrait yielded similar results to that of elf.tga; cropping the image ended up with a picture of my eye. Again, the scale didn’t appear to be small enough to crop the entire face, because it seems to have started at the correct position, but it stopped at cropping the eye.

Non-Smiling Group

findface on IMG_0011.tga with min_scale = 0.45, max_scale = 0.55, step = 0.05

Finding the faces on this image proved to work pretty well. I have no complaints with this one.

Smiling Group

findface on IMG_0012.tga with min_scale = 0.45, max_scale = 0.55, step = 0.05

As seen in this attempt, my overlap detection does not work correctly, so it did not get the face of the man on the left. After experimenting a bit, however, it will eventually detect his face if I include a few more faces to detect.

Class Picture

findface on IMG_0002.tga with min_scale = 0.45, max_scale = 0.55, step = 0.05

Apparently, Dave gets all the love in this shot. Again, my overlap detection doesn’t work correctly, so all the faces are focused on one part; namely, Dave’s papers. One of the faces did manage to detect Dave’s actual face, though.

Part 3: Verifying Faces

I tested verifyface with MSEs of 500, 750, and 1000, and 1250. The MSE of 1000 was the most effective, because there was a more substantial increase in accuracy between 750 and 1000 than with 1000 and 1250. There was not really much difference between using 1000 and 1250. The false negative rate ended up being about 12.5%, and the false positive rate ended up being around 20-25%.