PDF version of this entire document
The adjustment of the code continues based on observation of tough cases. And based on early numbers, the results do not improve much by extending the range of search (to more parts of the face) or increasing the number of vertices to 2,700. The reason is errors. The experiments were aborted within hours because the numbers were unsatisfactory. It is currently being explored just how much of the face surface can be used without causing trouble, just adding entropy. The error is basically a purely technical one which relates to FMM and triangulation.
The general strategy is currently to run false pairs on one 8-core server and true pairs on another, expanding the size of the sample and providing output as soon as any is made available, at the very least for supervision. For more comprehensive comparative experiments that help guide the development, it may be handy to have remote access to more machines with spare CPU cycles, e.g. machines at the lab that are unused overnight and have MATLAB installed. This would speed things up by parallelising experiments (they are serialised at this stage). Results generally continue to improve based on trials (and errors) that either show improvement or degradation. The currently ongoing experiment expands the scope explored on the surface and it also fixed the penalty terms based on careful observation of previous experiments. It is ideal never to have to use penalty, but for practical purpose it is necessary (although it is a hack).
Expanding the scope of surface matching leads to a higher frequency of errors, which in turn seem to exacerbate performance, based on an overnight experiment. Some of the remaining errors are due to poor ICP, which fails to properly align some basic structures like nose and eyes, in which case, the method basically does what it's intended to do. The problem is - at least sometimes - registration prior to geodesic analysis. This can probably be corrected without much difficulty.
Roy Schestowitz 2012-01-08
