Eigenfaces

Computer Vision Project 4: by Sellakumaran Kanagarathnam

Recognize Faces

Eigenfaces is a basic face detection and recognition program. By training the program on 32 images of my classmates, eigenfaces generated the following images for use in face detection and recognition.

Average Face

Top 10 Eigenfaces

Using a userbase generated from these 10 eigenfaces, my program correctly identified 20 smiling photos of the original 32 people used to construct the eigenfaces. This is a recognition accuracy of 62.5%. Below is a graph of the number of eigenfaces used and the resulting accuracy.

The benefit of adding more eigen faces was steep in the very beginning. It quickly reached the average peak at 8 eigen faces.  The peak benefit was at 12 eigen faces. 12 eigen faces seem to give the best accuracy, but looks like 8-12 would be a good number to settle at (again based only on this experiment).

There were also several consistent recognition errors.  Certain faces were never correctly identified. But interestingly, the correct face was within 2-4 best faces.

 

For example, s23.tga was never correctly identified .

In all except one case it was identified as

In almost all the cases, the second best picture was the correct picture for the above failure.

 Again s27.tga was never correctly identified:

In all except one case it was identified as

In almost all the cases, the second, third or fourth best picture was the correct picture for the above failure.

Find Faces

I had good success with cropping images given a picture with one person. But face detection in a group was good for the test picture, but was less than 100% for another.

min_scale = 0.25; max_scale = 0.55; step = 0.01; crop

 

min_scale = 0.25; max_scale = 0.55; step = 0.01; crop

 

min_scale = 0.25; max_scale = 0.55; step = 0.01; crop

 

 

min_scale = 0.5; max_scale = 0.70; step = 0.01; mark

This next image was less than optimal; there were 4 faces but my program & range only found 3 of them. Interesting note is that I was able to successfully crop the faces in individual portraits. The key is to find the right scales. For portraits, .25 to .55 seems to work fine for all my 3 samples. But the range of .5 to .7 does not work always for group pictures. I had to try a differnt range for the second picture. But I could not find a range where all four faces were detected. I had another range .4, .47, .05 where the first face was detected, but not the second one.

min_scale = 0.4; max_scale = 0.5; step = 0.01; mark

---

Extra Credits

1) Completed speedup of Eigen Vector calculation. The results were taken using the speedup version function. I still have the old function commented out in faces.cpp
2) I also completed verifyFace, a function that allows users to check an image against a userbase of known faces to determine if the face is who the user believes it to be. MSE tolerance of 2000 gets an accuracy of 62% and 3500 MSE tolerance gives 96% accuracy.

Tolerance MSE	% Accuracy
500	6
750	28
1000	40
1250	46
1500	53
1750	59
2000	62
2500	75
3000	87
3500	96