International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
* 
yft 
& 4. 
i 
"A\ 
Æ*l 
c. Patches from SPOT using REGSEG 
d. Patches from SPOT using OPTISEG 
Figure 6. Segmentation algorithm applied to SAR and SPOT. 
a. Patches from SAR using MUM 
b. Patches from SAR using RWSEG 
4. POLYGON MATCHING 
The polygon matching method used in PAIRS and 
ARCHANGEL are based on the work of Abbasi-Dezfouli and 
Freeman (1994) and developed by Dowman et al. (1996). We 
describe here a new method which has similarities with these 
earlier methods. The method is summarised in Figure 7. 
In previous work (Dowman et al. 1996, Dowman, 1998) chain 
coding has been used. This technique is not used here, as studies 
showed that it did not add to the success of results achieved 
from the techniques in Figure 7. 
The use of a cost function for image matching was described by 
both Morgado and Dowman (1997) and Abbasi-Dezfouli and 
Freeman (1994). However, both of these authors found that 
there are some problems associated with its implementation. A 
general point was that on its own it will not differentiate all the 
good matches from the bad ones. This is because incorrect 
matches sometimes give a low value for the cost function, and 
correct matches sometimes give high values. Therefore, more 
information has to be used to improve the matching results, such 
as using a chain code method to describe the shapes of the 
patches, or using the relative geometry of the matched patches 
with respect to each other in order to highlight poor matches. 
Although it was shown that both of these methods worked in the 
examples cited by the authors, they would not necessarily work 
here. The use of the chain code method was ruled out above. 
The problem with using the relative geometry of matched 
patches is that there may not be enough good matches to ensure 
a reliable result could be determined, or too many bad but 
geometrically similar matches could easily lead to an incorrect 
result. Therefore, there is a question of robustness associated 
with these two methods of improving the match results.