In contrast to other post-filtering strategies, this occlusion
filtering process is incorporated into the matching propaga-
tion. Particularly, this occlusion filtering method not only
helps to remove mismatches from occluded areas during
matching, but also provides clues for corrected matching,
thus leading to improved matching in occlusion areas.
Figure 4 shows an example of matching results obtained
with/without the occlusion filtering. Some points around the
buffer zones of the building roofs were matched incorrectly be-
cause of their similar intensities. These mismatches were elim-
inated by applying the occlusion filtering, and the matched
results coincided well with the building roof boundaries.
Segment-Adaptive Similarity Measurement
In image matching, the matching cost is normally exploited to
compare the relationships or similarities of the corresponding
pixels and thus determine a correlation coefficient. However,
the correlation assumes equal depths for all pixels within a
local window. Notably, this assumption violates situations
with occlusion problems caused by depth discontinuities.
Figure 5 illustrates two types of typical decision conflicts
regarding similarity measurement that occur during the
search for matches around building boundaries. In this case,
the correlation window must be split into two parts. One type
of decision conflict is a “border defect”, wherein parts
a
and
b
correspond to the two correlations
CC
a
and
CC
b
. The concept
that part
a
owns the higher correlation coefficient is easily
confirmed. However, the similarity measure of part
b
is lower
due to a mis-correspondence. As a result, the overall correla-
tion may not exceed the predefined threshold. This may lead
to missing of correct matches near the borders of the area. The
other type of decision conflict is “border blurring”. In Figure
5, the right-hand image (after splitting the correlation window
into two parts) contains three parts. However, parts
b
1
and
b
2
exhibit a mis-correspondence, leading to a higher similarity
for part
a
than for parts
b
1
and
b
2
. As a result, the overall cor-
relation may exceed the predefined threshold, causing exten-
sion of the roof beyond the building and blurred border.
Inspired by the segmentation-based occlusion filtering, a
segment-adaptive similarity measurement method is intro-
duced. Figure 6 illustrates the basic concept underlying the
method. Assuming the grey region in Figure 6a represents the
Figure 4. Example of matching results obtained with/without occlusion filtering.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2018
139
