obtained from the first
SIFT
run resulted in a relatively higher
RMSe
compared to the other runs. However, even with 211
points a subpixel registration
RMSe
was obtained. As it was
expected, with increased number of matching points, the reg-
istration
RMSe
decreases. Surprisingly, even though the num-
ber of matching points is higher in the second and fourth
SIFT
runs, the registration
RMSe
is relatively higher compared to
the third run (Table 2). Unlike the other
SIFT
runs, the points
obtained from the third run are concentrated on the near-
vertical outcrop with almost no points from the background
minimizing the depth-of-field, which could be the reason for
relatively lower registration
RMSe
.
The precision of matching points under different geometric
transformations were evaluated based on the error calculated
as the Euclidean distance of each matching point in the warp
image to its predicted location. In addition, distribution of in-
dividual point errors was also evaluated to determine the most
feasible geometric transformation for each
SIFT
runs (Figures 3
and 4). Regardless of the transformation,
SIFT
points are spread
over the outcrop and do not show any pattern in error distribu-
tion, that is the individual point errors appear to be random
for both affine and first-order polynomial transformation. The
highest individual point error from different geometric trans-
formations using different
SIFT
runs are also tabulated in Table
2. The accuracy of homologous points were further evaluated
Figure 3. Distribution of individual point error from (A) the first, (B) the second, (C) the third, and (D) the fourth runs in
affine transformation.
Figure 4. Distribution of individual point error from (A) the first, (B) the second, (C) the third, and (D) the fourth runs in first-
order polynomial transformation.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
December 2018
785
