September 2019 Full - page 681

Especially, the correspondences extracted by the proposed
deep descriptor PPD are nearly three times those of the SIFT
and BRISK descriptors. Figure 12 shows the correspondences
extracted using different combinations of detectors and
descriptors in v_graffiti . The correct matches are linked with
the green line. The combination of the SIFT detector and the
proposed PPD descriptor has the top performance in the image
pair i_leuven with illumination variation among all other
combinations. The handcrafted descriptor BRISK is appar-
ently better than the deep descriptors PPD and TFeat when the
interest points are detected using the BRISK detector. When the
SIFT detector was adopted to extractthe interest points, the
proposed deep descriptor PPD is better than the SIFT descriptor
and the deep descriptor TFeat. Figure 13 shows the corre-
spondences extracted using different combinations of detec-
tors and descriptors in i_leuven.
Remote sensing images: Satellite image with illumination
variation: Four satellite images acquired over the city of Paris,
France, on different dates ranging from October 21, 2007, to
February 26, 2019, downloaded from Google Earth are used
to show the effectiveness of the proposed descriptor under
illumination change. The satellite images with a resolution of
1644×944 pixels are group into three image pairs to show the
matches. Table 7 shows the basic information of the key points
using different feature detection algorithms, including BRISK
and SIFT, in the three image pairs. Although all descriptors can
find enough matches in three multi-temporal image pairs, the
proposed PPD descriptor is always better than SIFT, BRISK , and
TFeat as seen in the number of correspondences in Table 8. Fig-
ures 14, 15, and 16 show the correspondences in three multi-
temporal image pairs using different combinations of hand-
crafted feature detection algorithms and learned descriptors.
Aerial image with viewpoint variation: Three oblique
image pairs acquired over the city of Zurich, Switzerland,
with a medium-format camera (Leica RCD 30 Oblique Penta)
are used to demonstrate the effectiveness of the proposed
descriptor under viewpoint variation. We resize the original
images, free of lens distortion, with 9000×6732 to 3000×2244
Table 6. The basic information of the patches in the two image data set. KeyPoint1-B and KeyPoint2-B are the numbers of interest
points extracted from an image pair using the BRISK algorithm, while KeyPoint1-S and KeyPoint2-S use the SIFT algorithm.
Data Set Variation Resolution KeyPoints1-B KeyPoints2-B KeyPoints1-S KeyPoints2-S
i_leuven Illumination 921×614 719 2944 997 3297
v_graffiti Viewpoint 800×640 3524 5204 2678 3922
Table 7. The basic information of the four different acquired
satellite images.
Image 1 Image 2 Image 3 Image 4
Acquisition
Date
21 October
2007
16 April
2014
18 May
2018
26 February
2019
Resolution 1644×944 1644×944 1644×944 1644×944
BRISK 88 721 68 565 46 804 72 398
SIFT 42 392 41 241 36 284 35 900
Figure 11. The number of correspondences after outlier
removal using RANSAC for different detectors and
descriptors in data sets i_leuven and v_graffiti.
Table 8. The number of correspondences of the three image pairs using different combinations of detector and descriptors. PPD
is the proposed learned deep descriptor. Bold numbers indicate that the corresponding descriptors have the top performance.
BRISK BRISK+TFeat BRISK+PPD SIFT SIFT+TFeat SIFT+PPD
Image Pair1 (1–2) 399 194 810 418 209 539
Image Pair2 (1–3) 438 251 735 420 185 467
Image Pair3 (1–4) 305 111 437 239 102 263
Figure 12. Feature point matching results of image pair v_graffiti after outlier removal using RANSAC through different
combinations of different detectors and descriptors.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING September 2019 681