Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 90

poses without explicitly establishing correspondences between model features and
parts of the image.
Methods Based on Monocular Image Data A probabilistic approach to si-
multaneous pose estimation and object recognition is proposed by Niemann and
Hornegger ( 2001 ). However, they only regard the problem of 2D-2D pose esti-
mation. They follow a statistical approach to object recognition, while localisa-
tion is performed based on estimating the corresponding parameters. An object
model is represented by the position-dependent and typically multimodal proba-
bility densities of the pixel grey values in an image that displays an object of a
certain class. Relying on empirical data, the number of the components of these
multimodal distributions is determined by vector quantisation, while the param-
eters of the distribution are obtained with the expectation-maximisation algo-
rithm. Hence, pose estimation is performed based on a maximum likelihood esti-
mate.
In the monocular system of Kölzow and Ellenrieder ( 2003 ), a unified statisti-
cal approach for the integration of several local features, such as edges or tex-
tures, is employed. The utilised object model consists of CAD data complemented
by the local features, resulting in an 'operator model'. The six pose parameters
of a rigid object are then obtained by adaptation of the visible model features to
the image. The accuracy of the obtained pose estimation results is discussed in
Sect. 6.1 .
Other pose estimation algorithms rely on geometric (edges) and on intensity (sur-
face radiance) information. In this context, Nayar and Bolle ( 1996 ) introduce an ob-
ject representation based on reflectance ratios, which is used for object recognition
using monocular greyscale images. Pose estimation is performed relying on the re-
flectance ratio representation and a three-dimensional object model, thus taking into
account physical properties of the object surface in addition to purely geometric
information.
Another technique which relies on the simultaneous extraction of edge and shad-
ing information for 2D-3D pose estimation is the appearance-based approach pro-
posed by Nomura et al. ( 1996 ), who utilise synthetic edge and intensity images
generated based on an object model. A nonlinear optimisation procedure based
on a comparison between the observed and the synthetic images yields the pose
parameters. The accuracy of the obtained pose estimation results is discussed in
Sect. 6.1 .
The approach by Ando et al. ( 2005 ) is based on a set of grey value images of an
object with known associated three-dimensional poses. After reducing the dimen-
sion of the images by principal component analysis, the rotational pose parameters
of the object are learned directly from the image data using support vector regres-
sion.
The integrated 2D-3D pose estimation approach by Barrois and Wöhler ( 2007 ),
which in addition to edges and surface brightness also takes into account polarisa-
tion and defocus features, is discussed in Sects. 5.6 and 6.1 .
The three-dimensional pose estimation method of Lagger et al. ( 2008 ) is based
on an 'environment map', i.e. a direction-dependent map of the light intensity and