Computational Aspects: How Landmarks Can Be Observed, Stored, and Analysed - Landmarks: GIScience for Intelligent Services - page 141

Tabl e 5. 3 Different
weightings for different
aspects of salience, after [ 53 ]

Area Shape Color Visibility Marks

Day 0.11 0.15 0.37 0.26 0.12

Night 0.26 0.00 0.21 0.23 0.30

study, where participants had been asked to select the most prominent façade for
each intersection while viewing a 360 ı panorama photograph of that intersection.
In seven of the nine intersections, automatic selection corresponded to human
selection, showing the power of this formal model for landmark salience—or more
precisely façade salience.
The same authors also discussed approaches of adapting the parameters of the
model to specific contexts [ 53 ] . Again using an empirical study, they established
weighting factors for different aspects of the model when encountering façades
during day or night. People were asked to select the most prominent façade at an
intersection from a photograph (showing either a day or a night shot), and then
to rank different aspects according to their importance for making this selection.
Tab le 5.3 lists these factors. While visibility is important both during the day and at
night, shape does not seem to be used at all at night. Instead a façade's area becomes
much more important. The same holds for marks, especially if they are illuminated.
Elias [ 11 ] identified this challenge of data collection and parameter adaptation as
the weak spot of Raubal and Winter's salience model. She proposed to use existing
topographic or cadastral data sets and to run machine learning approaches to identify
potential landmark candidates. The attributes in these data sets can be used as feature
vectors describing the different geographic objects. These feature vectors may be
fed into classification or clustering algorithms, which will then identify 'outliers',
i.e., geographic objects that are not easily joined with other objects. Arguably, these
outliers stand out from their surrounding environment and, thus, can be considered
to be landmarks following Presson and Montello's definition [ 32 ] .
Depending on the chosen classification algorithm, using this approach may
require to normalize the data first. Attributes may need to be preprocessed such
that they are all on the same scale and use the same measurement type (ordinal,
nominal, etc.). This is to ensure that no single attribute dominates the discrimination
between objects. Elias focused on buildings as landmark candidates and proposed
the following attributes to use in the classification (see Table 5.4 ) . As can be seen,
these attributes either refer to land use or are derived from the geometry of buildings,
in other words information that can be expected from a cadastral data set. This
makes data collection easier, however, it also means that visual attractiveness can at
best only be implicitly inferred from non-visual attributes.
From the many possible approaches, Elias chose to test her approach using
ID3 [ 33 ] , a supervised classification algorithm, and Cobweb [ 13 ] , a hierarchical
clustering algorithm. While both approaches seem promising, ID3 has the advantage