Geodesic lenses

The entire image/data set has been stacked up inside loader functions for large experiments. Some special cases were then studied in order to work around them not by detecting but passing based on borderline scores. A bug in the penalties was found and corrected, even though these penalties depend on how the algorithm is varied (must be normalised wrt other variables).

I am working on nice ways of visualising the localised geodesic errors between pairs, densely. This ought to help indicate, e.g. using colour maps/contours, where two individuals differ (if at all), at the very least helping a human assessment which uses Euclidean (human-visible) by providing 'geodesic lenses'.

For debugging or general analysis that helps understand why the same imaged person can be intrinsically different across images, a tool was made to highlight localised differences such that for each pair of any objects (not just faces), the discrepancy will be visually identifiable and therefore possible to factor out, baaed on observation. Algorithms can be adjusted accordingly to avert false negatives. To be more effective, it will need to be remapped more like a compass and overlaid with some colours.

**Figure:** Two examples of easy matches from the remainder of the dataset (which was enrolled in its entirety into the experiment)

**Figure:** Example of geodesic differences map around the nose (to be improved)

The overlay of choice is a spiral where lines represent D sampled varying (increasing) distances away from the fiducial/key point, wherein degrees are represented in a way that can relate to the original images. Overlaying the images in a way which cannot obscure anything may require colour, though.

**Figure:** Example of a thin FMM spiral

Overlays of geodesic distance indicators are not easy to make visible, even by 'redifying' an intensity-scaled indicator of distance. At the moment, the output looks something like in:

**Figure:** Four small examples of distances spiral in isolation

**Figure:** 2 larger examples of distances derived from pairs of images of the same people

**Figure:** The distances spiral overlaid on the images it corresponds to (9th person in the set)

**Figure:** The distances spiral overlaid on the images it corresponds to (13th person in the set)

At the point where detection rates are improving it is usually the subtle localised errors (with very high values) that put the whole classifier at peril. Using these maps ought to help judge - on a case-by-case basis - the composition of the overall score. I am currently examining the image pairs with those charts apart in order to better understand what to tweak for improved performance, especially fewer false negatives. The points have been dilated someone to improve visibility, as shown in Figure $[*]$ .

**Figure:** The distances spiral with larger, clearer points