Johnathan Lyon :: CSE 455 Project 4 Artifact
Part 1: Testing Recognition with Cropped Class Images
Average Face
10 eigenfaces
Mismatch Example 1 : 13 eigenfaces - MSE: 53825.7
Mismatch Example 2 : 13 eigenfaces - MSE: 24513.2
Mismatch Example 3 : 25 eigenfaces - MSE: 22338.2
Mismatch Example 3 : 3 eigenfaces - MSE: 10315.5
1. DESCRIBE THE TRENDS YOU SEE IN YOUR PLOTS. DISCUSS THE TRADEOFFS; HOW MANY EIGENFACES SHOULD ONE USE? IS THERE A CLEAR ANSWER?
There appears to be a logarithmic shape to the results and that the results plateaus at 68% efficiency. The tradeoffs for using more faces is that you start to overspecify the face space, and adding more eigenfaces increase computation in the detection phase. However, you definitely want enough to get a good sense of the face space. For this set it seems that 11-13 faces would be sufficient (20 optimal but only for +4% change) however in general it is more difficult to say. This would depend on how characteristic the general various across all samples best describes the detection problem.
2. YOU LIKELY SAW SOME RECOGNITION ERRORS IN STEP 3; SHOW IMAGES OF A COUPLE. HOW REASONABLE WERE THE MISTAKES?
The mistakes became more reasonable as the number of eigenfaces was increased. Some of the mistakes it was obvious that there was a visual correlation, others, it seem to be possibly a well align feature issue. As you decreased the number of eigenfaces the problems were less apparent.
Part 2: Cropping and finding faces
elf cropped
(0.45-0.55 by 0.01 using 10 30x30 eigenfaces)
digital of self
digital of self cropped
(0.73-0.78 by 0.01)
digital of self marked
(0.73-0.78 by 0.01 for 2 faces using 10 30x30 eigenfaces)
group1 of 3 marked
(0.95-1.05 by 0.01 for 3 faces using 10 25x25 eigenfaces)
group4 of 3 marked
(0.45-0.55 by 0.01 for 3 faces using 10 30x30 eigenfaces)
class group2 marked
(1.0-1.5 by 0.01 for 10 faces using 10 30x30 eigenfaces)
class group2 marked
(0.85-1.5 by 0.01 for 12 faces using 10 30x30 eigenfaces)
group6 of 3 marked
(0.45-0.5 by 0.01 for 3 faces using 10 25x25 eigenfaces)
group6 of 3 marked
(0.45-0.5 by 0.01 for 3 faces using 10 30x30 eigenfaces)
1. WHAT MIN_SCALE, MAX_SCALE, AND SCALE STEP DID YOU USE FOR EACH IMAGE?
This is specified above each image.
2. DID YOU ATTEMPT TO FIND FACES RESULT IN ANY FALSE POSITIVES AND/OR FALSE NEGATIVES? DISCUSS EACH MISTAKE, AND WHY YOU THINK THEY MIGHT HAVE OCCURED.
For the digital self photo a false positive was detected. It was verified that the correct face was the 2nd best choice, however, the patch of hairy chest texture seemed to fool the face detection. This is probably due to a correspondence the hair pattern with the shading found in the face space, although it is difficult to say exactly as to the naked eye there seems to be little correspondence.
There were more false positives that were harder to discern for larger images. Some of this was fixed by altering the size of the eigenfaces used (see: group6 examples) or by increasing the number of faces to find (only marginally improved in group2). I did not have a skin color filter in place nor did I incorporate variance (as it seemed to make things worse in more cases) but I did see improvements by factoring in distance from the average face. The false positives ranged from generic texture spaces, to spaces that seems to have general shading and shape correlation to the average shading in the eigenfaces.
Part 3: Verify Face
1. WHAT MSE THRESHOLDS DID YOU TRY? WHICH ONE WORKED BEST? WHAT SEARCH METHOD DID YOU USE TO FIND IT?
I circumvented this issue by doing a direct analysis on the resultant MSEs. This is problematic as you can see from the distribution of cutoff values, that there is not a significant enough sample space to really understand the shape/behavior of the correct matches MSEs. By plotting the efficiency results, based on varying the cutoff values for each corresponding sorted correct match's MSE, we can choose a better threshold for this data set.
2. USING THE BEST MSE THRESHOLD, WHAT WAS THE FALSE NEGATIVE RATE? WHAT WAS THE FALSE POSITIVE RATE?
For an ideal threshold value of 105,000 (+/- 2,0000) I got a success rate of 81-83%, a false negative rate of 15-18% and a false positive rate of 15-18%.
Extra Credit
1. Morphing
Using 50x50 eig
Example using smiling images to generate the eigenfaces for smiling face morphing
Morph from 08.tga to 09.tga 0.0-1.0