According to the plot, the number of faces correctly recognized increases as the number of eigenfaces used increases. Hence, the data seems to suggest that we should use as many eigenfaces as possible. However, more eigenfaces used means a greater amount of storage is required. The computation time for reconstructing a face will also be longer. However, for the purposes of this test, I would recommend using 25 eigenfaces.
| Smiling | Non-smiling | Correct Answer Location | |
|---|---|---|---|
| 1 | |
|
8 |
| 2 | |
|
1 |
| 3 | |
|
19 |
As shown in the table, some of the correct answers appeared highly in the sorted results while others were very low in the list. Hence, the recognition technique can be said to be erratic at best when it fails to perform the correct recognition.
min_scale=0.1 max_scale=0.2 scale step=0.01
min_scale=0.45 max_scale=0.55 scale step=0.01
min_scale=0.45 max_scale=0.55 scale step=0.01
min_scale=1.0 max_scale=1.2 scale step=0.01
min_scale=0.35 max_scale=0.4 scale step=0.01
| Non-smiling student | Corresponding face | Corresponding face MSE | Closest smiling face of someone else | MSE |
|---|---|---|---|---|
 |
 |
9991.33 | |
13300 |
|
 |
187371 | |
16674.6 |
 |
 |
19977.6 |  |
29942.4 |
 |
 |
35471.3 | |
3920.81 |
 |
 |
36079.9 | |
46417.2 |
I tried the following MSE thresholds: 10000, 20000, 30000, 40000, 50000.
The best threshold was 30000. This gave a false negative rate of 50% and a false positive rate of 37.5%.
I used a linear search to determine the threshold that will give me the least sum of false negatives and false positives.
Morphing from image smiling_cropped\01.tga to image smiling_cropped\02.tga
Training images used
Result of findFace on a picture filled with dogs.
The following is a plot of the number of faces correctly recognized versus the number of eigenfaces used.
This plot uses a smaller image resolution of 10 X 10 pixels to generate the eigenfaces.
By comparing the 2 plots, we see that the smaller resolution yields a slightly better recognition as 1 additional image is correctly recognized for the greater number of eigenfaces used. Smaller resolutions might have better recognition because the eigenfaces will reflect the most important data instead of subtle aspects of the images that might be present in higher resolution images.
It would seem that a smaller resolution, in this case 10 X 10 pixels, would be optimal for a face.