Integrated Frameworks for Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 239

Fig. 5.21 Input intensity images for pose 1 ( left ) and pose 2 ( right ) of the rubber (cf. Table 5.6 )

Baumer industrial CCD camera of 1032 × 776 pixel image size, equipped with a

f 25 mm lens. The approximate distance to the object was 0 . 5 m. The coordi-

nate system was chosen such that the x and y axes correspond to the horizontal
and vertical image axis, respectively, while the z axis is parallel to the optical axis.
The scene was illuminated with a LED point light source located at a known po-
sition. The algorithm was initialised with four different poses, differing by several
degrees in the rotation angles and a few millimetres in translation. As a result of the
pose estimation we adopted the minimisation run yielding the lowest residual error
according to ( 5.46 ).
The reflectance function R I was determined with a goniometer by estimating
the parameters according to ( 5.22 ). At the same time we found that the polarisation
degree of the light reflected from the surface is so small that it cannot be reliably
determined. Hence, the input data for pose estimation are limited to intensity, edges,
and depth.
For our evaluation, we attached the rubber with its lateral surface to the goniome-
ter table and oriented it in two different poses relative to the camera. The angular
difference between the two poses is only a few degrees (cf. Fig. 5.21 ). For the de-
termination of the ground truth, we replaced the rubber for each pose by a chequer-
board of known geometry. The chequerboard was attached to the goniometer table,
and its pose was estimated using the rig finder algorithm described by Krüger et
al. ( 2004 ), which is based on a bundle adjustment approach for camera calibration
purposes. Due to the simple cuboid shape of the rubber, the chequerboard pattern
could be aligned at high accuracy into the same direction as the lateral surfaces of
the rubber, such that the chequerboard pose could be assumed to be identical with
the pose of the rubber.
The results of this experiment are shown in Table 5.6 . The deviations between
the measured and the true pose parameters are only a few tenths of a degree for
the rotation angles and a few tenths of a millimetre for the lateral translations. The
translation in z is determined at an accuracy of about 4 mm (which is about an order
of magnitude lower than the lateral accuracy) or 1 %. This is a reasonable result,
given that only monocular image data are available.

=