correlation coefficients in this region. However, when using
the segment-adaptive similarity correlation, the pixels within
the roof region receive significantly greater weights in the
similarity measurement. Accordingly, the pixels on the ground
have a weaker influence on the similarity measurement.
Dense Matching Propagation
Based on the feature matching results from the previous steps,
a dense matching propagation is carried out to obtain dense
matching results for subsequent 3D surface reconstruction.
The dense matching propagation includes the following two
strategies: local matching propagation and regional match-
ing propagation. Figure 8 provides a framework for the dense
matching propagation.
First, a seed match list comprising all vertices of the
triangulations from the previous feature matching is gen-
erated. The points in the seed list are sorted according to
their correlation coefficients in feature matching. The dense
matching propagation begins with the most reliable match in
the seed list and propagates the matching of this point to its
neighborhood using the local matching propagation strategy.
The newly matched points are then inserted into the seed
match list based on their correlation coefficients. Subsequent
matches in the seed list are then selected for matching propa-
gation until all points have been processed. Some regions in
the images may remain unmatched (i.e., cavities) after local
matching propagation and are subjected to regional matching
propagation. The aforementioned disparity planes, segment
constraint, and segment-adaptive similarity measurement are
used for both local and regional matching propagation. The
final output is a set of dense matching results from which
photogrammetric point clouds and
DSMs
can be generated.
The local matching propagation follows the “best-first”
strategy used in popular dense matching methods (Lhuillier
and Quan, 2002; Megyesi and Chetverikov, 2004; Zhang and
Gruen, 2006). In this strategy, the most reliable seeds are
used to direct the matching propagation process. For a listed
seed located in a pre-defined segment, the disparities of the
24 connected neighboring pixels (a 5 × 5 neighborhood),
u
,
in one image will be interpolated from the corresponding
disparity plane. If the seed is located beyond any pre-defined
segment, the disparity of the seed itself will be assigned to its
neighboring pixels. The match candidates of the neighboring
pixels in the other image
u'
can then be estimated based on
the disparities. For each neighboring pixel
u
i
, a search range
(
u'
i
–
d
T
,
u'
i
+
d
T
) centered at
u'
i
along the epipolar line will
be examined; here,
d
T
is a pre-defined minimum searching
distance. The pixel with the highest correlation coefficient
within the search range is labeled as a match. Once a newly
matched pair of points is obtained, it is inserted into the seed
list according to the correlation coefficient. If all 24 neighbor-
ing pixels of a seed have been processed, the seed will be
removed from the list. After all of the listed seeds have been
processed, the local matching propagation is terminated.
Notably, the local matching propagation incorporates the seg-
ment constraint. In other words, if some of the 24 neighboring
pixels of a seed are beyond the segment, matching propa-
gation will stop at those pixels. The previously described
segment-adaptive similarity measurement is also used to
determine the correlation coefficients.
Local matching propagation has the following advantages.
First, its use guarantees robust and stable matching, as the
robust matches are propagated before others according to the
“best-first” strategy. Second, the disparity constraint facilitates
the matching of points in highly textured areas to those in
poorly textured areas. However, local matching propagation
occurs in immediate neighboring regions and may be cut off
by the coefficient threshold, possibly leading to the formation
Figure 8. Framework of the dense matching propagation.
142
March 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
