The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
196 
substantially in terms of translations, rotations, and scale 
changes. It should also be noted that a link (between nodes a 
and e) in the map network does not exist in the image network. 
This introduces inexact matching in the two networks, but only 
in their structure. 
Figure 4. Detected networks and their graphs 
Figure 5 illustrates the convergence of the global compatibility 
function under successive iterations until a maximum value is 
reached. The result using all three attributes is shown by the 
thinner curve (top) and its global compatibility increases faster 
and converges earlier than when using two attributes 
(connectivity and relative distance) only. The run time for this 
experiment is 1.2218 seconds (with two attributes) and 1.4821 
seconds (with all three attributes). 
Figure 5. Comparison under inexact matching 
The matching result is summarized in Table 1. It can be easily 
seen that all nodes were matched correctly despite differences 
in orientation (rotation, shift, and scales) between the two 
networks, or even differences in their actual structure (the 
presence of the a-e link). 
a 
b 
c 
d 
e 
f 
Matching 
result 
V 6 
V, 
V 2 
Vs 
v 4 
V 3 
Table 1. Matching result 
5.2 Test 2 
In this test, we examine the robustness of our approach in exact 
matching. M 2 in Figure 6 and M' in Figure 4 are two detected 
networks used in this experiment. It should be noted that M 2 has 
9 intersections and 13 edges, while M' only has 6 intersections 
and 7 edges. These two datasets vary not only in structures like 
the example in Test 1, but also in nodes of the graph. 
Figure 6. Detected network M 2 from the map 
In addition, as shown in Figure 7, there are four components in 
M 2 marked with colors that have same topological pattern as 
M'. Thus, topological attributes only would produce multiple 
results. 
Figure 7. Topologically similar components 
The convergence of the global compatibility is shown in Figure 
8. Similar as Figure 5, the result using all three attributes is 
shown by the red curve (top) and its global compatibility 
increases faster and converges earlier than when using two 
attributes only. In this test, the matching starts to converge 
after 30 iterations, slower than in Test 1 as we have relatively 
complex networks for matching. 
Figure 8. Global compatibility vs. iteration times 
The matching result is graphically described in Figure 9. 
Despite the topological similarity problems shown in Figure 7, 
intersections in M' are correctly mapped to M 2 .