constraint (marked as
WSC
and NSC, respectively) are shown
in Figure 8.
It is observed that less correct matches have been obtained
and the matching precision has dropped by more than 10
percent without using the seed matches to compute geometric
constraint. The main reason is that in order to improve the
efficiency of the
LDF
detector, the descriptor comparison range
in the support region distinctiveness factor computation is
reduced from the whole image to sub-region, and therefore it is
difficult to ensure that the support region of the detected fea-
ture point is distinctive in the whole image. When the exhaus-
tive search is performed without geometric constraint, similar
features from different sub-regions may produce false matches.
Evaluation of the Proposed
FIPS
Matching Strategy
In order to evaluate the proposed
FIPS
matching strategy, three
detectors are adopted to detect features in the correspondence
search area, i.e., the Harris detector, the exhaustive search
method (all pixels in the search area) and the proposed
FIPS
.
For all methods, features in the source image are detected by
using the proposed
LDF
detector, and the
USIFT
descriptors
are extracted. The
NNDR
method is used to find matches. The
test is based on the image pairs described in Table 3 and the
results of them are shown in Figure 9.
It is observed from Figure 9 that the exhaustive search
method and the
FIPS
-based method perform better than the
Harris based method both in terms of
NCM
and
MP
because
more pixels in the search area are checked. Comparatively,
the exhaustive search method produces slightly more correct
matches than the
FIPS
based method. However, the exhaustive
search method needs to calculate feature descriptors of all
pixels in the search area, resulting in very low time efficiency.
Compared with that, the feature descriptors of most of the
pixels in the search set of the
FIPS
method have already been
calculated in the feature point detection step. There is no
need to spend much time to calculate feature descriptors, and
the matching efficiency is very high. Therefore, the
FIPS
based
method is more practical by considering the overall perfor-
mance of matching result and algorithm efficiency.
Matching Methods Comparison
A brief description of the comparative methods is shown in
Table 5. The statistic results of our experiment based on the
automatic checking strategy are shown in Figure 10.
It is observed from Figure 10
that the proposed method outper-
forms the comparative methods in
terms of
NCM
and
MP
. There are two
main reasons: First, the proposed
LDF
detector finds more distinctive
features from repetitive patterns,
thus leading to higher matching
correctness.
LDF
detector serves
like a pre-filter ensuring most
of the input features easy to be
distinguished in repetitive texture
patterns. Second, the proposed
FIPS
is a set with denser features as
compared to other points detected
by traditional feature detectors.
Consequently, the
FIPS
based
matching strategy generates higher
NCM
and
MP
values. From the
results, we also observed that the
highest
MP
of the proposed method
is less than 80 percent. This is due
to the fact that the fitted geometric
transformation is only an approxi-
mation (e.g., assume the ground
plane is the flat), while cannot
accurately represent the physical
displacement resulted from the 3D
terrain relief.
The
GR-SURF
method achieves
the second best performance. The
algorithm utilizes the geo-refer-
encing information to construct
geometric constraint, which elimi-
nates false matches with large error
and finally improves the matching
precision (
MP
). However, the inputs
of the matching procedure are still
features with highly repetitive pat-
terns such that the improvement on
NCM
is limited.
The
GC
-
SIFT
,
GC
-
DAISY
, and
PG
-
DAISY
methods are the same type
of methods taking the advantage of
global information to distinguish
local repetitive patterns. More
specifically, the
PG
-
DAISY
method
Figure 7. Evaluation of the proposed
LDF
detector: (a)
NCM
values, and (b)
MP
values.
Figure 8. Evaluation of the influence of the seed matches constraint on the matching
performance: (a)
NCM
values, and (b)
MP
values.
Figure 9. Evaluation of the proposed
FIPS
based matching strategy: (a)
NCM
values, and
(b)
MP
values.
520
August 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
