PDF version of this entire document

Similarity Measure Variants

Several more approaches - variants of the original algorithm - were tested and later aborted as they did not show improvements. It does not seem as though realignments help in all cases because although they can resolve some problematic cases, they also ruin perfectly fine ones at times. What complicates things is the duration required to test variants, as even by running false pairs on one 8-core computational server and true pairs on another would typically take about an hour (for useful conclusions to be drawn).

Figure: The result of applying a faster calculation of similarity, as applied to the first person against 90 different pairs from 90 different people

We are thinking of testing a couple of new measures for alignment, among which are scale invariant and equi-affine invariant distances on surfaces. It could be interesting to see if those could boost somewhat the alignment process.