Geography Reference
In-Depth Information
4.3.1 Detection of Spatial and Spatio-Temporal Clusters
Clustering can serve as a primary tool for organizing the collection of photos into
groups. Among the possible methods used for spatial clustering are: grid based
(Girardin et al.
2008a
), density based (Andrienko et al.
2009
; Kisilevich et al.
2010a
) and hierarchical clustering (Zheng et al.
2009a
).
Clustering based on grids is data independent and does not take into consid-
eration the distribution of points. The number of cells should be known in advance
and many trial and errors are required to find the suitable number and size of the
cells. Density based clustering is based on the neighborhood density (min points)
and minimum distance between points using (usually Euclidean) distance func-
tions. The method produces an arbitrary number of clusters based on the selected
parameters. In general, variations of density based clustering can be applied where
the time component is taken into consideration (Andrienko and Andrienko
2009
).
For example, the spatio-temporal cluster will be formed if there are more than five
people that took photos within the range of 1 h and the distance between them is
no more than 100 m. Such an approach would create event-centered temporal
clusters.
Hierarchical clustering can be applied to form clusters at different scale levels.
Clusters on every level can be analyzed separately and different semantics can be
applied at different scales. For example, the tag that identifies the name of a city
can be assigned to the cluster on the city scale, while tags that identify names of
neighborhoods will be assigned to clusters at the neighborhood scale. Spatial
clusters can be produced by bounding the data with time limits. In this way, only
the data that falls into the time interval will be clustered whereas the clustering
algorithm will cluster points without taking the temporal aspect explicitly.
4.3.2 Text Analysis
Text analysis of titles and tags can be used for finding events that happen in a
cluster or in different parts of the world at the same time or at different times. For
this, the representative tags and titles can be obtained for several clusters and
matched for similarity. Examples of such events are New-Year celebrations that
take place at the same time in different parts of the country or the world. At the
global scale, difference in time zones should be taken into account by clustering
every region separately with adjusted time intervals.
4.3.3 Content-Based Analysis
Similarity between images in a cluster can facilitate finding different contexts. For
example a photo may not have title and tags, or its title is meaningless for analysis
(written in a language not known to the analyst or does not represent any event or
place). The visual similarity can be still found between other images in a cluster.
