te a
Tay
int.
| by
[he
ath
Ist
ion
of
rch
pel
ent
"try
nal
be
ery
ity
an
uch
Ahmed Elaksher
4.4 The Purdue Project
The Purdue images cover the campus area in West Lafayette, Indiana. The neat area from the corresponding images was
extracted and examined as illustrated in Figure 5.
Figure 5. Left and Right Images from Purdue Pair
This area contains small and large buildings, abundant walkways, and parking lots. There are also some occlusions,
shadows, and moving objects. The original photographs were nearly vertical and their nominal scale was 1:5000. They
were originally scanned into digital form with a resolution of 15 micrometers. Their aggregated version at level 2 (30
micrometers) was employed in this experiment, so one pixel represents approximately 0.3 m. in the object space.
The straight line, plateau, and spike features were extracted from both images. About 140 epipolar lines covering the
whole area were examined. There were approximately 50-70 extracted feature points from each epipolar line. However,
only 35-40 correspondences were found from each set. Some false alarms and missed detections of match pairs still
occurred. These errors are caused by the imperfect results from the feature extraction process, occlusions and other
factors. The 3D information computed from feature matching was used as constraint information in the signal matching
process.
Searching in the signal matching process was performed independently in two directions, north-south and east-west as
shown in Figure 6.
Figure 6. Orientation of the Vertical Planes Defining the Cost Matrices
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 271