PDF version of this entire document
There was a one-week period when several ideas were tested, including a mask that is spectral rather than geodesic. It was tested for potential improvements. Not much documentation was produced for it, due to technical limitations associated with distance from my main workstation.
We ended up spending a day building, testing, and running experiments with a spectral similarity measure for faces, but could not guarantee exciting results. The experiments from that day (scripted run over the weekend) indicated that there was merit in the approach tested, but in order for the ROC curve to look decent, a lot more work will be required.
