Computational Aspects: How Landmarks Can Be Observed, Stored, and Analysed - Landmarks: GIScience for Intelligent Services - page 143

Algorithm 5.1: Algorithm to generate a leveled landmark hierarchy, from [ 55 ] .
Data : l k is the landmark k on hierarchy level i ; L i is the set of all landmarks at level i ; L i k is

the set of all landmarks in the immediate neighborhood of landmark l k on level i .
1 forall the landmark l k at level i , l k 2 L i do

2 Compute the most salient landmark l k:max in its immediate neighborhood

L i k Dfl j jdist.l j l k / 1; l j 2 L i

g.
3 Add the most salient landmarks at level i as the set of landmarks at level i C1:
fl k:max jl k 2 L i

g fl i C r g.

4if jL i C 1

j >1j then

Compute the Voronoi partition and the Delaunay triangulation of L i C 1 5 and go back to

step 1 .
6else
7 Stop.

The emerging cells can be considered to be the influence region of a land-
mark [ 55 ] . As with most models, this strictly leveled approach to a landmark
hierarchy simplifies reality. It is well conceivable to have more subtle variations
in (perceived) landmark salience, which may lead to less strict hierarchies, or sub-
hierarchies within a single level.

5.2.2 Data Mining Salience

So far, we have discussed methods and techniques to identify landmark candidates
from geographic data. It is also conceivable to identify landmarks in less structured
geographic content, for example, from images or texts that depict or describe spatial
situations. In this case, methods from information retrieval, and geographic infor-
mation retrieval [ 20 ] more specifically, come into play. These usually encompass
some machine learning or clustering approaches to mine relevant information—here
landmark candidates—from unstructured or semi-structured sources.
Such approaches aim at finding relevant information in the selected data sources
where relevance depends on the question asked. For example, Zhang et al. [ 56 , 57 ]
extracted route directions, in particular their destinations, from web sites. They
achieved this by classifying elements of the text based on both the structure of
the text and the HTML tags, similarity of elements to prototypical patterns (e.g.,
'turn left/right at LOCATION'), the position of textual elements in the overall text,
and other parameters. Similarly, Drymonas and Pfoser [ 7 ] reconstruct tourist walks
from travel guides by extracting motion activities ('walk along...', 'turn left...')
together with the associated points of interest. These points of interest in turn can be
geocoded again, which allows for mapping the walks onto geographic data. While
none of these approaches searches for landmark candidates explicitly, some of the
methods would be applicable for such an aim as well.