PDF version of this entire document
Further experiments which look at Euclidean measurements upon geodesic circles (no GMDS) were able to smoothen the previous curve a bit and change in the range of circles improved recognition rates somewhat. This is not a substitute but a complement for GMDS - specifically for cases where GMDS does not provide a satisfactory classification (high uncertainty). This newer approach can be further refined although it takes overnight experiments to autonomously 'manufacture' a decent ROC curve that provides sufficient comparative insight. I shall place more markers on the image to apply FMM to as it ought to amplify the signal and cancel out some of the noise (e.g. beards and other acquisition errors). No fiducial points have been used thus far, obviating the need for any human intervention in this process. No texture data is being used, either, just the raw surface.
|
