275 
3. EXPERIMENT 
3.1 Data-Sets 
There are 5 groups' data for experiment, including along-track 
stereo image pairs and across-track stereo image pairs. 
® SCENE01 n SCENE02 : Data are stereo SPOT-HRG 
images, covering the area in France with 5m 
resolution^ 12000 column * 12000 row). 
(D SCENE03: The SPOT-HRG image covers the area in 
France with 10m resolution^ 6000 column * 6000 row). 
(3) Stereo Radarsat images cover the Erhai area in 
China ,with the sizes (8949 column * 8598 row )and 
(12865 column *12679 row). 
© Stereo pairs of SPOT imagery and Radarsat imagery, 
which cover the Hulu Island in China, with the sizes are 
(6000 column * 6000 row) and (7971 column *8910 row). 
(5) Stereo IKONOS images cover Beijing City in China, with 
lm resolution, and the sizes are ( 7853column * 19043 
row)and ( 7648 column * 19043 row). 
3.2 Experiment I 
The first experiment with the RPC model was performed on the 
satellite images deriving from various sensors to confirm the 
accuracy of forward and inverse transformation is up to 
standard. 
Select points which are uniform distribution on all over the left 
image to project to the ground by the forward RPC Model, and 
then project the ground points to the image back to obtain some 
points as result. Table 1 compares the coordinate of image 
points obtained by calculation with selected initially. The 
variance is presented as an accuracy index. Table I shows that 
no big differences exist in terms of points’ coordinates. This 
outcome indicates that the accuracy of forward and inverse 
RFCs are precision reached. 
Image 
Maximal 
difference in 
line 
Minimal 
difference in 
line 
RMSE in line 
Maximal 
difference in 
sample 
Minimal 
difference in 
sample 
RMSE in 
sample 
SPOT-HRG 
6.42E-02 
8.79E-09 
1.30E-02 
1.25E-02 
6.31 E-10 
2.30E-03 
RADARSAT-1 
2.39E-03 
-3.92E-10 
1.30E-05 
1.43E-03 
1.84E-09 
1.33E-05 
IKONOS 
2.22E-06 
4.96E-11 
1.39E-06 
7.68E-06 
1.82E-11 
1.76E-06 
QUICKBIRD 
-5.03E-05 
2.79E-10 
9.91E-06 
3.94E-05 
5.98E-10 
7.80E-06 
SPOT 
3.77E-04 
1.22E-08 
2.42E-03 
5.02E-04 
9.27E-09 
9.88E-05 
Table. 1 Accuracy of the forward and inverse transformation (pixels) 
3.3 Experiment II 
The experiment on the satellite images for epipolar resampling 
is carried out. 
Before the epipolar resampling process, a two-step pre 
processing is applied for the original images. First, the 
geometrical rectification is performed on the satellite stereo 
images using The Satellite Imagery Block Adjustment system 
software. Second, the correction coefficients of image 
coordinates deriving from the first step are used for affine 
transformation of image points. 
Therefore, bias-corrected images can be used to test the 
epipolar resampling algorithm. 
Do the same experiment to the five pairs of stereo-images. First, 
from the left image, a stochastic point is extracted to project to 
the ground according to the given elevation level. Change the 
height of the corresponding object point twice along the light 
ray connecting the perspective centre and project the object 
points onto the right image. There are three image points on the 
right image deriving from the projection. Fit these three points 
to be the right epipolar line. 
And then, choose 11 points evenly on the right epipolar line to 
project to the ground, the same as before, change the height of 
the corresponding object point and project the object points 
onto the left image. There are 11 image points on the left image 
deriving from the projection. Fit these 11 points to be the left 
epipolar line. 
Figs (1) to (7) show the generated conjugate epipolar lines and 
display them on the stereo-pairs. Fig (8) is a flow diagram 
showing the performance of the method of epipolar resampling. 
left image (SCENE01) right image (SCENE02) 
Figure 1. SPOT-HRG scenes 
left image (SCENE03) right image (SCENE01)