that 
e 
| with 
faster 
n the 
ne 
he 
ation 
tch is 
y. 
nd the 
izes of 
been 
large 
bal 
in the 
at the 
oise. 
18 
ach 
| 
  
th the 
nce 
rea to 
ing 
han 
find a 
> the 
nd this 
nding 
same 
indow of 
10sen. 
e 
r each 
8 
Jnew = 8 
Zi=0 Wi 
where w; is the correlation, and x; and y; are the 
coordinates of pixel ; in the window. The new 
corresponding point with the coordinates of (x 
should give sub-pixel accuracy. 
od 
4. Results 
Tests have been performed on sections of SPOT images 
and on some close range images obtained from the 
Internet. Examples of the results can be seen in Figure 
1 and in the following table. 
  
  
  
  
  
Image set A B C 
Corresponding points found for | 360 284 706 
correct patches 
Corresponding points found 0 164 
from incorrect patches 
Correct points 242 205 024 
Overall success rate (fraction 67 72 72 
of identified points that are 
correct, %) 
Success rate from correct 67 72 89 
patches (%) 
  
As before, images A and B come from a single SPOT 
pair collected 45 months apart, and image C is from a 
SPOT pair captured six months apart. With the B 
images, in all cases where the patches were incorrectly 
matched, no points were identified by the search 
procedure as having sufficient correlation to match, so 
the incorrect patches did not reduce to the final success 
rate of the point matching procedure. With the C image 
pair, some of the points bordering the incorrect patches 
had a high enough correlation to register as matches in 
our point matching procedure. These unfortunately 
reduced the overall success rate. For our procedure to 
work completely unsupervised, we will need to 
overcome the patch matching errors for point matching 
to proceed. 
In cases where a patch was correctly matched, there 
were often differences in the identified patch outline. 
Our search procedure was often successful in 
accommodating the patch outline error, and found 
correct point matches well away from the identified 
patch boundary. This can be seen on the right side of 
the patch in Figure 1b, and also in the lower left corner 
of the patch. 
This gives us confidence that the method is sound in 
principle. When the method was applied to two pairs of 
close range photographs, success rates of 91% and 85% 
were achieved from patches, all of which were in the 
correct general area, but again were areas of gradual 
colour change rather than clearly identified features. 
The present method of proceeding from patches to 
points attempts to find a match for every boundary 
pixel of the patches in the first image. In practice, this 
105 
  
Fig. 1(a) Section of left sub-image from C image pair 
  
Fig. 1(b) Section of right sub-image from C image pair 
Figure 1: Result of point matching with one patch, 
showing the success in finding matching points, 
despite an incorrect patch boundary having been 
identified. The image colours have been exaggerated 
by histogram equalization; the actual gray values lie 
in quite a narrow range, and the overall intensity of 
the two images is quite different. 
will give too dense a set of matched points for most 
purposes. It also presents problems in areas where 
there is little colour variation along the boundary. A 
better approach will be to confine the search in some 
way, such as by identifying boundary points with high 
interest and finding their match, before proceeding with 
other boundary points or points away from the patch. 
4. Conclusions 
We have demonstrated the principle of operation of our 
proposed method and have had some encouraging 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996