GMDS-PCA: Project Plan

Dr. Roy Schestowitz

Summary: Estimated milestone towards comparable
results to state of the art


-------------------------------------------


After about 300 hours of work, including work on technical 
documentation and learning of the associated methods, the target 
design is generally in place. Its form is still crude at places 
and therefore it is worth considering what was previously done 
better, eying potential reuse of external or peer code, even 
though there are implementational overlaps. One can identify the 
following as items requiring completion in order for results to 
improve further, based on observation of improper identifications 
(more of those appear, especially as false pairs are added and 
get misclassified, reducing recognition rate down to ~93%):

1. consistently ordering PCA for sampling of distances/spatial 
  points, maybe finding the right balance between both

2. cheek sampling/completion (a slider for that added)

3. mustache as an issue motivates heightening the binary mask to 
  never account for facial hair

4. partial matching needed for cases where hair hides the eye 
  region

5. cropping more consistently to further assist GMDS

6. less smoothing, more hole filling instead, especially around 
  the eyes

7. better ICP for initialisation of points

The above requires little actual work and more testing, including 
the building of large models that depend on choice of parameters 
and need to be rebuilt upon changes to the algorithm. A lot of 
this would be passive, meaning that it requires letting the 
computational servers (now 2 are readily available, not one) run 
overnight, whereas actual work -- that which is being measured -- 
should last just dozens of hours. Face-to-face consultation with 
people at the lab has helped considerably when it comes to 
eliminating poorer solutions and exploring what previously 
worked, first in 2005 when GMDS was applied to and tested on face 
data, then around 2009 when recognition rate as measured on large 
datasets reached 97%.

The problem of face recognition is a thoroughly explored one for 
reasons that are simple to grasp. The effectiveness of some 
methods in this already-crowded space is due to use of the rich 
structural information that is also accurate at the photometric 
side, permitting accurate measurements to be taken (both in 2- 
and 3-D). The strength of GMDS lies within the fact that many 
geodesic distances can be measured quickly, e.g. using FMM with 
friendly hardware architecture. This, in turn, gives many sample 
points on the surface. PCA can autonomously assist the weighting 
of the different distances, just as GMDS autonomously finds 
approximations of analogous points. By forcing GMDS to latch onto 
features that are more easily identifiable, competitive 
performance can probably be attained. 

One suggested path to explore would be to try the finer quality 
database from Texas. Generally speaking, one of the profound 
advantages (or 'selling points') of GMDS is that it is generic, 
thus it need not depend on a priori knowledge or markup about the 
problem domain. To measure the merit of GMDS based on face data 
is like judging a Swiss army knife based solely on the sharpness 
of its blades. Its versatility and problem-agnosticism ought to 
accentuate its real power, hinged upon sophistication and not ad 
hoc or brute force methods. While perfectly acceptable given the 
ubiquity of the application of face recognition, this might not 
be the ideal domain on which to use GMDS. The intrinsic 
appearance of faces is simple enough for a human observer to 
interpret accurately, whereas it's flexible tissue and underlying 
morphology where humans lose the ability to discern one from 
another.