Previous: Projection onto Face SpaceJAFFE Results
For the JAFFE database, our algorithm recognized a new photo of the person 65% to 75% of the time. We defined recognition as successfully matching a test image to a picture of the same person from the training set, even though the two images had different expressions. Recognition rate increased dramatically as we increased the number of eigenfaces, but stabilized after four eigenfaces (that is, our recognition rate did not increase by using more than four eigenfaces). The recognition rate also increased when we used more test images per person in the training set. Similarly, the recognition rate increased when we used a greater percentage of the dataset in the training set (with pictures chosen randomly from the dataset).
 |
Rice University Field Test Results
For the Rice University database, our algorithm recognized a new photo of the person 65% to 75% of the time. Again, recognition rate increased dramatically as we increased the number of eigenfaces. In this case, the recognition rate stabilized after six eigenfaces. The recognition rate required more eigenfaces to stabilize because the Rice database was more diverse, including people of both genders and many different ethnicities. Therefore, more eigenfaces are needed for an accurate representation of the Rice population. Again, the recognition rate increased when we used more test images per person in the training set. The recognition rate also increased when we used a greater percentage of the dataset in the training set.
 |
The resulting final algorithm was able to determine, at the four rates shown in the previous two figures, whether a test image was a non-face or face and also a match to known face, finding the closest match to a known face (as shown below).
 |
