FMM-based Dissimilarity

In existing (ongoing) experiments, rather than vary the geodesic boundaries, the locations of the points get altered, under the assumption that this can provide a greater source of variability, covering a greater extent of the surface being probed (in isolation for separability of regions). This is not stochastic yet, but it can be made so.

The results are interesting so far (no mis-detections), but more of them are required to draw some meaningful conclusions. GMDS might not be ideal for measuring FMM-dependent similarity, so composing a substitute or complement for this task might make sense, improving it one step of complexity at a time (assessing what improves it and what does not). Ultimately, perhaps a problem-specific or similarity-optimised method can be devised as a substitute rather than a fallback for GMDS and/or PCA (where scale and thus speed/memory are an issue). The sensitivity of GMDS was at times also a weakness, matching things that oughtn't be matched without a penalty large enough.

**Figure:** Interim results (70 images) show 95% recognition rate with FMM-only (no GMDS) utility, but this tends to degrade as more difficult images are presented. Two good recognisers (classifiers), one of which is a Euclidean-geodesic hybrid, might give pretty good and mutually-independent results without using texture or fiducial points.

Various images that GMDS deals with just fine are not handled as easily by this other method I gradually refine (a hybrid of FMM and a level sets-inspired technique), so they can correct one another and make a better joint recogniser. One problematic pair, just for the sake of an example, is shown in Figure $[*]$ , which is basically detected as almost belonging to the same person (it is actually on the margin of uncertainty), so the new method ought to be made more sensitive and less permissive. Currently, the results it yields can be seen in Figure $[*]$ and Figure $[*]$ .

**Figure:** An example of two images from two different people, which nonetheless the FMM-based recogniser cannot quite detect as being different

**Figure:** An FMM-based recogniser results in nearly 90% recognition rate now (without GMDS)

**Figure:** The FMM recogniser ROC curve after increasing the number of true pairs