Average Face Plus Top 10 Eigenfaces
For part 1 of this experiment, I generated 10 eigenfaces, at 100x100 as well as the average face (on the left). I didn't use eigenfaces this large throughout these experiments, but I include them here for visualization purposes.
|
|
|
|
|
|
|
|
|
|
|
For parts 2 and 3 of this experiment, I investigated the effect of the number of eigenfaces used on the accuracy of face recognition. The chart below shows the results using between 1 and 31 eigenfaces. The eigenfaces were all 25x25 and there were 32 faces total that I attempted to recognize.
This graph appears to indicate that, for best results, it is probably best to pick roughly 10 or more eigenfaces. There didn't seem to be a significant penalty, with regards to efficiency, with choosing more eigenfaces. For each eigenface count used in the experiment, I could generate the eigenfaces, form the user base, and run the recognizer on all 32 faces in under 5 seconds. Now, you will note that the accuracy actually decreased, from 62.5% to 59.375% when the number of eigenfaces was between 15 and 19, but then increased again when the number of eigenfaces went beyond 19. My only explanation for this is that the 17th eigenface must push the one additional misclassified face further towards the wrong face in face space but then when we add additional eigenfaces it pushes it back towards the correct face. Below I show the face that was incorrectly recognized, the correct face that it should have matched with, the face that it incorrectly matched with, and the 17th eigenface that caused the mismatch. You will see that the reason why this eigenface has this effect is not obvious.
 |
 |
 |
 |
In most of the cases, the misclassifications were understandable. The "smiling" faces that were radically different in pose from their non-smiling counterparts were not recognized with any number of eigenfaces (see below for a couple of examples). They often did not appear high in the sorted list of matches either. I noted that variations in face angle often lead to a misrecognition. The mismatches that did appear high in the sorted order of results were often those that were recognized after adding additional eigenfaces.
 |
 |
 |
 |