Computational Aspects: How Landmarks Can Be Observed, Stored, and Analysed - Landmarks: GIScience for Intelligent Services - page 145

a b

c d

Fig. 5.5 A toy example for hierarchical clustering of POIs. Initially, every POI is its own
cluster ( a ). In a second step, the closest two POIs form a new cluster. The new cluster center—
the centroid—is shown in black , the original POIs in dark grey ( b ). This process is repeated until
there is only a single cluster left ( c, d )

former by the latter gives a number that indicates how unique a given n-gram is for a
cluster because higher numbers indicate that this n-gram appears frequently within
a cluster, but only sparsely outside it. In addition, Mummidi and Krumm also define
a minimum size for a valid cluster, and measure 'term purity', i.e., the fraction of
descriptions within a cluster that contain a given n-gram.
Others have used tags associated with Flickr photographs to find appropriate
labels for places or to delineate city cores [ 18 , 34 ] , using similar ideas, but different
techniques to those just presented.
Next, we will have a look at approaches that aim at identifying prototypical
and/or prominent views for specific locations from photograph collections. For
example, Kennedy and Naaman [ 23 ] as well as Zheng et al. [ 58 ] use unsupervised
learning techniques to find canonical (prototypical) views for given clusters of
photographs that (likely) show the same geographic object. They use Flick r 1 or
Picasa 2 and Panoramio 3 as data sources, respectively, exploiting tags, geographic
position, and the images themselves.

1 http://www.flickr.com , last visited 8/1/2014
2 https://picasaweb.google.com , last visited 8/1/2014
3 http://www.panoramio.com , last visited 8/1/2014