Conclusions: What Is Known and What Is Still Challenging About Landmarks - Landmarks: GIScience for Intelligent Services - page 203

feasible approach uses (spatial) reasoning mechanisms to identify landmarks
based on, among others, a geographic object's location. Likely, in the end a
combination of these two approaches will prove most successful.
Independent of the chosen approach, having the right data available might
ultimately be the greatest challenge. In Chaps. 5 and 6 , we have seen that today's
approaches rely on detailed data being available for all potential landmark
candidates across an environment. This is an unrealistic demand, even for an
approach such as the one by Duckham et al. [ 5 ] that only needs type information
for POIs.
2. Context: No system has yet a grasp of context in any way comparable to a person.
Face-to-face communication includes a multitude of often subconscious and
automated considerations regarding the partner's physiognomy, mood, or other
stimulations of visual, auditory, tactile or other senses. People are highly skilled
in non-verbal communication, and constantly make inferences about context
without recognizable effort. Machines have a very limited sensory input, which
is one reason why they have a hard time with non-verbal communication—think
about reading your communication partner's mood in a video conference call
with poor connection.
But even already the verbal (or graphical) communication is challenging for
a computer. Understanding the semantics of language and making correct infer-
ences is one of the big research challenges in artificial intelligence. Language is
under-specified and hence ambiguous. We have seen examples of this challenge,
such as the ambiguity of the reference frame for the interpretation of spatial
relationships. Does “Melbourne Cup” mean a race carried out in Melbourne, or
is it a proper name? Does “left of the school” mean left in the allocentric system
of the school, or in the egocentric system of the person providing this utterance
while standing in front of the school? If somebody is looking for a kindergarten
near their home, is that near the same as in “the café near the library”? As a
final example, and referring to the mentioned contrast sets [ 18 ] , is the spatial
granularity of the references to the Eiffel Tower the same in “Let's meet on the
observation deck of the Eiffel Tower” and “I have visited Notre Dame, Louvre
and the Eiffel Tower”? A person would interpret these ambiguous messages
by considering the given communication context and some default reasoning
[ 10 , 19 ] , while the computer would stick with term recognition, and perhaps some
stochastic interpretations from machine learning.
In terms of providing knowledge for our intelligent geospatial system we envision a
combination of several different approaches using traditional data sets, data mining
of 'neo-geography' sources, and user-generated content platforms.
Google Street View or Nokia City Scene have increasing coverage—such
panorama views of the environment may be the closest a system may ever get to
an embodied experience. While automatically identifying signage and shop labels
in these images is an impressive achievement, it is only the first step, and easier
than, say, automatically identifying churches or bus stops. Even more challenging
is identifying the 'outstanding' building in a given scene. Remember the grey