Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 91

colour incident on the object surface. An image sequence showing the object is ac-
quired, where a three-dimensional model of the object is available. For each image,
the diffuse reflection component is separated from the specular component, where
the latter is then used to infer the corresponding environment map. The pose pa-
rameters are refined by minimising an error measure which consists of a weighted
sum of the negative normalised cross-correlation coefficients between the pixel grey
values of the first and the current image and the squared differences between the cor-
responding environment maps.
Chang et al. ( 2009 ) propose two appearance-based methods for estimating the
three-dimensional pose of objects with specularly reflecting surfaces based on
CAD data. A mirror-like reflectance behaviour of the object surface is assumed.
The first approach is based on the determination of the position and appearance
of specular reflections in the image by rendering and a subsequent comparison
with the observed image. The second technique relies on the 'specular flow',
i.e. the optical flow associated with specularly reflecting surface parts. The ac-
curacy of the obtained three-dimensional pose estimation results is discussed in
Sect. 6.1 .
Methods Based on Multiocular Image Data Classical monocular pose estima-
tion approaches have in common that they are not able to estimate the distance to
an object at high accuracy, since the only available depth information is the scale of
a known object in the resulting image, and the appearance of the object in the im-
age is not very sensitive to small depth variations. In comparison, for a convergent
stereo setup with a baseline similar to the object distance, for geometrical reasons
a depth accuracy of the same order as the lateral translational accuracy is obtain-
able. Accordingly, a variety of three-dimensional pose estimation methods relying
on multiple images of the scene have been proposed.
The system of Bachler et al. ( 1999 ) performs a pose estimation of industrial parts
based on a pair of stereo images in the context of taking the parts out of a bin with
an industrial robot. In a first step, planes are detected in the scene relying on the
projection of structured light. After isolating an object based on the plane detection
result, the rotation angle around the optical axis is determined. Furthermore, the
translational pose parameters parallel to the image plane are estimated using a CAD
model, while the distance to the object is assumed to be known. The accuracy of the
approach is discussed in Sect. 6.1 .
A fast tracking algorithm for estimating the pose of an automotive part us-
ing a pair of stereo images is presented by Yoon et al. ( 2003 ). This method
is regarded further in the application scenario of industrial quality inspection in
Sect. 6.1 .
Rosenhahn et al. ( 2003 ) introduce a method for appearance-based three-
dimensional pose estimation from several images, regarding objects characterised
by free-form surfaces modelled in terms of Fourier descriptors, thus providing a
computationally favourable approximate representation on large spatial scales. The
pose parameters are determined in the framework of conformal geometric algebra.
The optimisation of the pose parameters is performed based on the appearance of