![](bg2.png)
due to the conceptual shift from using traditional 2D
CAD drawings towards 3D models within the build-
ing and construction industry sector. The arising chal-
lenges for architectural 3D collections have been ad-
dressed by a number of projects, e.g. MACE and
PROBADO. MACE is a former EU-funded project
(2006-2009) that aims to connect and improve acces-
sibility of various repositories of architectural knowl-
edge and enrich their contents with metadata (Meta-
data for Architectural Contents in Europe, 2006).
PROBADO was a project funded by the German Re-
search Foundation (2006-2011). Amongst others, its
goal was to integrate 3D architectural models into the
librarian process chain, starting with acquisition over
indexing up to presentation/delivery (Berndt et al.,
2010). Figure 1 shows the resulting visualization of
a key-word search. Current activities like the DU-
RAARK (Durable Architectural Knowledge) project,
which just started in 2013, concentrate on the long-
term preservation aspects of architectural 3D content
(Durable Architectural Knowledge, 2013).
The appealing presentation of retrieval results of
the main field of applications for our presented ap-
proach – an automatic camera animation, which is
generated during the indexing process of a repository.
2 RELATED WORK
Drucker and Zeltzer described a framework for ex-
ploring intelligent camera controls in a 3D virtual en-
vironment. They did not allow arbitrary 3D mod-
els, but their algorithm is constraint to indoor scenes
where solely a path from one room to another room is
computed in a virtual museum application (Drucker
and Zeltzer, 1994). They used the A* algorithm based
on (Hart et al., 1968) and the Manhattan (L
1
) distance
as a metric.
Argelaguet and Andujar (2010) focused on the is-
sue of computing the ideal speed of the animation
depending on the coherency of consecutive frames
to provide non-fatiguing, informative, interestingness
and concise animations. However, the camera path to
be used was given and not computed automatically.
Hence, their work allows arbitrary scenes, but the ac-
tual camera path is defined manually.
Santos and Duarte generated a graph of valid po-
sitions in a building using an adapted version of the
probabilistic road map generation algorithm proposed
by (Salomon et al., 2003). They used this graph to
guide a user through an unknown building using an
appropriate animation of the camera (dos Santos and
Duarte, 2011).
Ahmed and Eades (2005) described an algorithm
to compute a smooth transition of the camera from
one target node in a graph to another. However, they
mainly addressed Focus+Context issues involved in
navigating large graphs in 3D.
Stoev and Straßer (2002) proposed a method to
compute “good” views of a given data set not only
based on the projected area, but also on the depth of
a specific view. These views are used to generate a
camera path out of it. They tested their approach for
historical data sets like terrain models.
One group of automated camera control ap-
proaches are the visual servoing approaches, which
are a type of reactive approaches, since they react on
changes. These approaches are computationally effi-
cient and thus also suitable for highly dynamic envi-
ronments. E.g., such approaches can be used to follow
a person in a computer game, while avoiding obsta-
cles and occlusions (Espiau et al., 1992) and (Courty
and Marchand, 2001).
Another group of automated camera control ap-
proaches are the pure optimization based approaches.
There, shot properties are expressed as objectives,
which are maximized with classical determinis-
tic (gradient based, Gauss-Seidl, etc.) and non-
deterministic (such as genetic algorithms (Oliver
et al., 1999), Monte Carlo based, etc.) optimization
algorithms.
Christie et al. (2005) extended the work of (Oliver
et al., 1999) for a dynamic camera, where the con-
trol points of a quadratic spline curve are optimized.
However, they used a fixed look-at point and also
known start and end-position of the camera path.
Viola et al. (2006) and (Sokolov et al., 2006) tried
to characterize cognitive aspects such as scene under-
standing and attention. The former work showed how
to compute a set of characteristic views of a scene,
where the latter focuses more on scene understand-
ing. They generated automatic camera paths too. To
do so, they used a heuristic optimization technique
that relies on a local neighborhood search, where they
try to attract the camera to unexplored areas of the
scene.
Representative of the class of constrained based
automated camera control approaches are (Jardillier
and Langu
´
enou, 1998) and (Christie et al., 2002). In
the approach of Jardillier et al. the path of the camera
is created by defining a set of properties on the desired
shot. Christie et al. proposed enhancements in terms
of more expressive camera path constraints.
Since pure optimization techniques and also pure
constraint based approaches have some drawbacks,
some works try to combine these two techniques; see
(Christie and Olivier, 2009).
A definition of viewpoint quality with respect to
GRAPP2014-InternationalConferenceonComputerGraphicsTheoryandApplications
336