An important limitation to Sela's real-time implementation is its use of
binary edge magnitudes. Thus, the gradient map must be thresholded. This
process reduces the sensitivity of the computations since the values of
(contrast) are limited to 1 bit of dynamic range.
In Equation , the symmetry lines are computed by
considering the set of edges in
.
The latter can be
visualized as a set of annular regions or rings centered around point p (see
Figure
). Only pairs of edges
from the edge map falling within the ring will be used to compute the symmetry
lines at that point. Furthermore, edge magnitudes must be attenuated if their
normals are misaligned with the normals of the annular sampling region. The
normals of the annular sampling region are merely the normals of its contour
as shown in Figure
. The contribution of a
cocircular edge is proportional to its magnitude projected onto the normals of
the annular sampling region [37]. However, Sela deals exclusively
with binary values for edge magnitude. Thus, he totally discards the
contribution of edges whose phase angle is not within a range of the normal of
the annular sampling region they fall into. His computation only considers
edges whose phase is
degrees from the normal of the annular
sampling region instead of gradually attenuating the contribution of
misaligned edges. A value of
=6 degrees is typical. Thus, only edges
somewhat parallel or anti-parallel with the normal of the annular sampling
region they fall into are used in computing cocircularity.
Figure
depicts the filtering performed to discard
(instead of attenuate) the misaligned edges.
![]() |
In Sela [42], the dynamic range of r is limited to 3 bits as is the
dynamic range of .
In other words, only 8 different symmetry
orientations at 8 different scales (sampling rings) are computed (see
Figure
). Furthermore, the phase description of the
cocircular edges is also limited to one of 8 possible values. The reduction of
dynamic range enables Sela to utilize pre-computed lookup tables to obtain the
symmetry and interest maps instead of performing detailed calculations during
the execution of the transform.