Integrated Frameworks for Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 235

by an OpenGL-based renderer using model information about the object provided
by CAD data. The comparison provides an error measure which is minimised by
an iterative optimisation algorithm. Although all six degrees of freedom are esti-
mated, the described approach requires only a monocular camera, circumventing
disadvantages of multiocular camera systems such as the need for extrinsic cam-
era calibration. The framework is open for the inclusion of independently acquired
depth data. A first evaluation is performed based on a simple but realistic example
object.
Classical monocular pose estimation approaches have in common that they are
not able to estimate the distance to the object at reasonable accuracy, since the only
available information is the scale of a known object in the resulting image. Scale
information yields no accurate results, since for small distance variations the object
scale does not change significantly. In comparison, for a convergent stereo setup
with a baseline similar to the object distance, for geometrical reasons a depth ac-
curacy of the same order as the lateral translational accuracy is obtainable. For this
reason, a variety of three-dimensional pose estimation methods relying on multiple
images of the scene have been proposed (cf. Sect. 2 for an overview). However,
from a practical point of view, using a monocular camera system is often favourable
(cf. Sect. 6.1 ), although a high pose estimation accuracy may be required, e.g. to
detect subtle deviations between the true and desired object poses. In this section
we therefore regard a monocular configuration.
5.6.1 Photometric and Polarimetric Information
The first source of information we exploit is the intensity reflected from the object
surface. For this purpose, we make use of the two-component specular reflectance
function inspired by Nayar et al. ( 1991 ) and defined by ( 5.22 ). The unknown surface
albedo ρ is estimated by the optimisation algorithm described below. Although we
regard objects of uniform surface albedo in our experiments, our framework would
in principle allow us to render and investigate objects with a textured surface by
using texture mapping in combination with an estimation of the factor ρ . The other
parameters of the reflectance function are determined empirically as described in
Sect. 3.2.1 , regarding a sample of the corresponding surface material attached to a
goniometer.
The determined parameters of the reflectance function and a CAD model of the
object are used to generate a synthetic image of the observed scene. For this pur-
pose, an OpenGL-based renderer has been implemented. The surface orientation is
required for each point of the object surface to compute a reflectance value accord-
ing to ( 5.22 ), but OpenGL does not directly provide this information. Hence, the
technique developed by Decaudin ( 1996 ) is used to calculate the surface normal
for every pixel based on three rendered colour images obtained by a red, a green,
and a blue virtual light source appropriately distributed in space. Afterwards, the
reflectance function ( 5.22 ) is used to compute the predicted intensity for each pixel.