GMDS on Smaller Face Parts

**Figure:** A look at an alternative mask which focuses on the nose and inner eye only

**Figure:** Recognition results based on the mask from $[*]$ with GMDS

**Figure:** A nose-only mask, which omits areas with potential of facial hair (the examples at the centre and the left are not related)

**Figure:** The performance attained by applying GMDS just to the nose region

**Figure:** Example of the effect of ICP-induced rotation on the Voronoi cells

**Figure:** Return to the old mask with additional rotation, which does not yield better results than those at the region of 92%-98% recognition rate

It seems that we have most of the components to have a perfect system, but maybe the MDS implementation is not by the book, as the results are not as one would have expected. We are aware of half a dozen deficiencies and will address each one of them in turn. It is also apparent that we need to take into account special cases that recur. Figures $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , and $[*]$ show the results of some further debugging and gradual tweaking.

**Figure:** Cheek inclusion gradually staged in for understanding of its impact on recognition performance (geodesics and PCA)

**Figure:** The stress map corresponding to the new binary mask (with 150 points for FMM)

**Figure:** Stress map and the corresponding faces (looking from beneath the nose) from which it is derived

**Figure:** An example of a false pair (different people) and a cleaned up stress map showing some interesting patterns

**Figure:** Another example of a false pair and the results of GMDS

**Figure:** Example of a bug found in the program, leading to massively false correspondence upon the same person

**Figure:** An example of acceptable matching between two poses of the same person

**Figure:** Another example of a bug found (and resolved) after it had proven problematic to recognition rates

**Figure:** A look at the problem associated with narrow faces that lead to incompatible sampling

Overnight, large experiments were run for 6 hours, flagging quite clearly all the cases that remain problematic and need closer attention as the false recognitions generalise to other examples of their kind. Some mistakes are caused by bugs in the code, especially in situations like special cases or bad data.

For images with only minor expressions we still hover at over 95% recognition rate. The problematic case are ones where the variation is great (between semesters for example) and there is partial matching in need.

We're working our way up, gradually improving performance by identifying edge cases and addressing them with some more sophisticated and problem-specific code which in turn generalises to more images exhibiting the same problem. Some examples of the progress are visualised in figures $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ , $[*]$ $[*]$ , $[*]$ , and $[*]$ .

**Figure:** General program settings used for the subsequent experiments

**Figure:** Results of a large-scale test after previous bugfixes

**Figure:** Example of a correspondence problem in a pair of images (one image on the left, another on the right). The top 4 images show the correspondence after the bugfix for one pair and the bottom 4 show the outcome of applying a fix to another pair.

**Figure:** Example of a problematic pair where hair obstruction and nose position compared to the forehead caused an issue which is now properly addressed

**Figure:** Example of a pair where the side of the face got sampled, leading to serious issues (top) before they got resolved (bottom)

**Figure:** Comparison between images of the same person, where the height of the nose relative to the cropping is causing issues

**Figure:** The images corresponding to the above example (same person, different positions)

**Figure:** Another example of a problematic example where the score borders on being seen as ``no match'' even though it is

**Figure:** Some recognition results from the above experiments, with denser sample on the right where the cheeks were also remove to test their impact on performance (little impact)

**Figure:** Smaller-scale and large-scale (right) experiments that look at how applying the methods only to the training set (many identical faces clustered together) changes the above results. It does not affect them much.