Travis McCoy: EigenFaces

Testing recognition with cropped class images

Procedures:

1. Cropped, non-smiling eigenfaces:

Average:

Eigenface0:

Eigenface1:

Eigenface2:

Eigenface3:

Eigenface4:

Eigenface5:

Eigenface6:

Eigenface7:

Eigenface8:

Eigenface9:

3. The number of faces recognized based on # of faces considered: (x axis = #used y-axis = #recognized)

Question 1: Based on the data, one could argue that two ranges of values for number of faces to be considered would be valid, depending on one’s preference for performance over accuracy. 9 faces appears to have the best trade-off between speed and accuracy, offering good accuracy for a minimal number of faces. At the same time, 18 faces gives the same accuracy as 25 faces, but uses 7 fewer faces. Consequently, there is no clear answer, and thus depending on the needs of the user, 9 or 18 would both be good choices.

Question 2: image 2 has a hard time recognizing the smiling version of the picture. The two are shown below:

While the correct result is listed as the 8^th strongest match, the pictures aren’t dramatically different in angle or lighting, so we should reasonably expect the accuracy to be better.

In contrast, image 11 doesn’t match as the first result, but is recognized as the second. The images are shown below:

The two images are very close to each other, and as we would expect, they are ranked closely, though not as each other’s top match.

Cropping and finding faces

Due to time constraints, I was unable to successfully complete this feature by the deadline. Please note that all other 6 TODO methods in the project are present and working properly.

Verify Face

Procedure:

For the first six smiling and non-smiling images, they had the following MSE’s

	1	2	3	4	5	6
1	9395	205716	66967	13144	73356	282957
2	35541	193812	70681	35228	43806	204327
3	153856	64697	20657	163729	265130	135602
4	46887	143308	36932	35599	127625	182140
5	86473	333305	153801	74181	40944	327688
6	564335	263465	269796	517585	597838	76881

Rather than search for the best threshold, we can plot out the MSE for each image pair and compute the best threshold. From the data, we see that that a mean squared error threshold of approximately 41000 generates the best results, balancing between finding all matches and limiting false positives. With such a threshold, we have a false negative rate of 33%, and a false positive rate of 13%. It’s worth noting that to get a better false negative rate, we would have to increase the threshold to 77000, dramatically increasing the false positive.