PDF version of this entire document

Smoothing

Following some further investigation it seemed reasonable to try smoothing of areas like the eyes, where local inconsistencies got GMDS preoccupied. So far the results suggest perfect separation (set size is about 30 and growing).

A problematic borderline case is shown in Figure and also the sorts of surfaces (in 3-D, see Figure ) where GMDS oddly enough failed, despite trying different resolutions and random seeds.

Figure: Problematic real pair (same person) where GMDS works but poorly so

Figure: A 3-D representation of a pair of images from the same person

A GMDS-based face recognition task, with smoothed surfaces where the resolution is increased for accuracy and for improved performance, still works rather well (room remains for improvement). In the following experiment only one image was problematic, only slightly bordering the threshold because of pose variation on the face of it (still needs further investigation, see Figure ). There was only one case where GMDS failed and the reason is yet unknown (Figure ). The ROC curve is in Figure .

Figure: GMDS failing to work as expected

Figure: A problematic pair which is seen as too different to quality as a match

Figure: ROC curve based on the smoothed surfaces variant of the algorithm

A kernel/window 13 pixels across, moving average (horizontal and vertical). The 2-D Gaussian filter is another option, but although it's in the program, it is not in the GUI yet, so this has not been attempted (there is a lot more that can be tested). Back in July-August it could be demonstrated that by smoothing the surfaces, errors could be reduced somewhat.