PDF version of this entire document
Shown in the image grabbed for Figure ![[*]](/IMG/latex/crossref.png)
are the results one gets from applying GMDS with default parameters
to datasets comprising a variety of expressions and no constraints
on scope, except exclusion of non-frontal face parts, including the
neck, hair, ears, etc. As expected all along, the performance takes
a noticeable hit. It might be interesting to see what putting/piping
the distances through PCA will do to overall performance, at the very
least on a relative scale. It might also be interesting to see what
performance we get by just looking at the eyes and nose in isolation,
perhaps LDAing them having used photometric data for segmentation
and then applied GMDS several times. If we had decided to limit the
measurement of geodesic distances to only particular segments, this
would be simple to implement.
With some new results from overnight experiments, it seems unlikely that adding the cheeks will improve performance much, to say the least (it is too inconsistent there). The current line of work looks at piece-wise GMDS, wherein facial features are taken in isolation to see the discriminative power of each.
Regarding eye-sockets., it is worth thinking about measuring local distances in a Euclidean fashion and longer ones, the geodesic way. This might be a best of both worlds approach, assuming of course that best detectors use the former method and the latter can complement it.
Alternatively, we can just be switching to Euclidean at regions of suspected peculiarities (missing parts) and large depth variations assuming the feature detector could isolate the feature points accurately. This is probably where using texture would be helpful. There must be existing implementations for segmentation of the face based on intensity data alone. The results with all cases really look bad thus far.
