Computational Aspects: How Landmarks Can Be Observed, Stored, and Analysed - Landmarks: GIScience for Intelligent Services - page 147

Fig. 5.6 The canonical views for several of Europe's most photographed cities; from http://www.
cs.cornell.edu/w8/~crandall/maps/map-europe.png (last visited 15/11/2013). Figure used with

permission by the author

The mean-shift technique is a statistical approach for estimating parameters
of an underlying probability distribution. It is non-parametric, which is
advantageous because no input parameters have to be guessed or arbitrarily
set. More specifically, mean-shift provides the modes. The mode is the value
that appears most often in a distribution.
In the case of finding the most popular landmarks, the underlying probabil-
ity distribution of where in a space people take photos is unobservable, i.e., it
is impossible to find a functional description of that distribution. Still, mean-
shift allows for finding the modes, which correspond to those places where a
lot of people take a photo, i.e., places that are interesting or important.

For the identified landmarks, Crandall et al. look for canonical views, i.e., typical
photographs. This combines clustering with graph construction. Images are tested
for similarity using SIFT (Scale Invariant Feature Transform; see [ 28 ] ) interest
points. A graph is constructed with each photo as a node and each edge between
a pair of nodes is weighted with the similarity between the two photos. The graph
is partitioned to find clusters of similar photos. For each cluster the node (photo)
with the largest weighted degree is chosen as the canonical view for this cluster.
Figure 5.6 shows an example of canonical views for some of Europe's most popular
cities in form of a map.