Trained the system using the cropped, nonsmiling face images, then tested the system's ability to recognize the same faces from the cropped, smiling images.  Varied the number of eigenfaces from 1 to 33. 

The maximum accuracy was about 70%, which only required about 5 eigenfaces. More than 5 eigenfaces didn't improve the results, as shown in the graph. The greater the change in expression, the less likely the system was able to recognize that person, which seemed to be a bigger problem than the number of eigenfaces. 

The optimum number of eigenfaces is about 5, there was no benefit to using more than that.

 

The neutral face on the left was consistently mistaken for the smiling face on the right regardless of the number of eigenfaces. 

The average face (left) and the top 4 eigenfaces.

Trained the system using the cropped, nonsmiling face images, then tested the system's ability to find faces in the group photos. 

The best strategy was to manually examine the approximate size of the faces in the group photo, then use a scale factor so that would be in the same approximate size as the eigenfaces. For example, if most faces were about 50 pixels wide in the group photo, and the eigenfaces were 25 pixels wide, use minScale=0.3, maxScale=0.7, step=0.1. 

The system is pretty good at identifying faces, but also has a lot of false positives (the background, clothing, etc.). Reducing the number of faces to find didn't help because some of the false positives had better scores than real faces. 

When using the original photos the cropped photos were taken from, better results were achieved, which makes sense since these were the same photos the eigenfaces were computed from.