PDF version of this entire document

False Positive - a Dilemma

Whilst in the process of programming a way around false positives (different people detected as identical) the following image was produced to provide insight into the process. For reasons of speed/pace of progress, not many such figures were previously prepared.

Shown in the images (see Figure [*]) are 4 problematic image pairs in a test set where one pair is wrongly said to be similar (by GMDS alone) and 3 others are close to the margin, which still causes false positives to be encountered. On the surface, it should be easy to tell them apart, but based on a geodetic test around the eyes and nose alone, it is hard to draw the line between identical and non-identical (need leniency as well as stringency). Measurement of distances between landmarks would work well, but it would miss the whole point of using an intrinsic geometry-based method, merely emulating popular/mainstream methods instead.

Figure: 4 problematic image pairs
Image false-positive

We now hope latch onto similar parts of faces. We could marry the classics and GMDS, extracting the best from both worlds.

This might defeat the purpose of the original plan/intent, but then again, it would not be research if we knew all along that GMDS can yield 99% recognition rate without - say - LDA applied to Euclidean-geodesics conjunction. I shall come up with a way of intertwining those two logically such that they are not treated as entirely separate and independent (but texture will still be discarded, only the surface will be used).

Roy Schestowitz 2012-01-08