In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009 
197 
Figure 9: Evolution of the precision against the number of 
images retrieved. 
5.3 Results on dataset 2 
An example of results using the brute voting is shown on 
fig Fig. 10. As we can see, none of the top images are 
relevant. The same occurs with the angle differences re 
finement. 
Figure 10: Example of first images retrieved using the k- 
NN voting for the second subset. 
The RANSAC refinement (Fig. 11) is able to retrieve two 
relevant images within the first five images, which means 
that irrelevant matches have been well filtered out. 
Like we did for the first subset, we compute the mean best 
rank shown in table 2. We were not able to compare with 
linear k-NN search due to the time taken by this method. 
The first observation is that none of the methods is able to 
retrieve even one relevant image within the top ten, which 
means that the methods are not able to give satisfying re 
sults from the users point of view. Nevertheless, the geo 
Figure 11: Example of first images retrieved using the 
Ransac refinement for the second subset. 
Method 
mean best rank 
Image matching 
80.67 
Brute vote 
98.80 
Ransac 
34.40 
Angle Differences 
59.10 
Table 2: Mean best rank for the second dataset. 
metric consistency step (either Ransac or the Angle Differ 
ences) provides a nice improvement. 
Figure 12: Evolution of the number of relevant images 
against the number of images retrieved. 
The evolution of the number of relevant images is shown 
on Fig. 12. As we can see, all methods are almost equiv 
alent, with the Ransac strategy being a little better for the 
last 20 images of the top 50. 
The precision is shown on Fig. 13, and is very low for 
all methods. The best result is obtained for the Ransac