The implementation incorporates a filter component which carves out
shape residue based on intuition and not a clear description of a
method elsewhere, e.g. in the IJCV paper (there is none specified
in the paper, just the observation that it's essential). They appear
to be using a hybrid of tricks to achieve this, so we start with threshold
that is permissive enough to accept natural variation. Next, a special
case to remove border differences will be implemented. Figure
shows a real example as it is being processed in phase 1.
In order to improve models that are being built automatically, i.e.
without human intervention, effective removal of noise and outliers
is done with spatial thresholds (depth and boundaries perpendicular
to the camera). There are more examples in Figure .
These are examples of outlier stripping in the shape residue, without
further cleaning of differences near face borders.
Roy Schestowitz 2012-01-08