![](https://csdnimg.cn/release/download_crawler_static/12673206/bg14.jpg)
1 Introduction 3
Since SRA requires autonomy [37, p. 35] from future robots, they will have to
be able to learn models of scenes on their own and in a multitude of domains. The
subsequent requirements regarding the generality and flexibility of a suitable model
for scene recognition are directly related to this observation.
1. A uniform representation of spatial relations that is sufficiently generic to describe
each type of relation but still captures the details in its variations.
2. Freedom in choosing which pairs of objects within a scene category are intercon-
nected by spatial relations.
A mobile household robot, which we take as an example, is going to face both
missing objects and clutter when trying to match a learnt scene classifier to its per-
cepts. Hence follows another requirement for scene classification:
3. Robustness against missing objects and clutter.
Existing work on probabilistic methods for part-based object recognition suffers
from severe limitations to all these requirements. As an alternative, we propose to
derive scene recognition from another method from part-based recognition that is
called Implicit Shape Models (ISMs)
2
[30], a Generalized Hough Transform [8]
variant.
In the usual household, scenes will not be visible as a whole from a single point
of view. In order to nevertheless gather evidence about existing scenes, robots will
have to integrate successive estimates about the presence of objects while freely
traversing their environment. Thus, a scene category representation should be favored
that fulfills the following requirements:
4. Independence from the viewpoint from which a scene is perceived.
5. Low time consumption, since scene recognition is executed repeatedly during
evidence-gathering.
Requirement 4 is best met with a scene category model in which both object poses
and spatial relations are specified in six-degrees-of-freedom (6-DoF). In literature,
such three-dimensional models are usually limited to modeling object positions with-
out orientations. In addition, we expect scene category representations to consider
uncertainties in spatial relations. Since modeling uncertainties in 6-DoF with para-
metrical distributions is tedious, a non-parametrical approach as the ISMs is more
appropriate for us. However, it is an open issue how to adapt their r epresentation
and the algorithms operating on them in order to model scenes in full 6-DoF while
maintaining efficiency. The s ame holds for requirement 2, as ISMs in the field of
part-based object recognition are only able to represent relations between a single
so-called reference part and all other parts of the object instead of relating arbitrary
combinations of parts of an object [21, p. 70].
Household scenes like in Fig. 1.2 can, e.g., be described as a single global scene
or as a combination of local scenes of presumably different categories. Since local
2
Even though the inventors of the Implicit Shape Models present a probabilistic motivation for their
approach in [30], we do not regard it as a probabilistic approach in a strict sense.