Integrated Frameworks for Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 230

three-dimensional point cloud (step 4) yields a dense three-dimensional surface pro-
file which displays a similar small-scale structure as the initial three-dimensional
profile but a lower large-scale slope. Repeating steps 1-4 during subsequent itera-
tions and taking into account the intensity information of camera 2 further refines
the three-dimensional surface profile.
The specular stereo method explicitly models the appearance of the surface in the
two cameras based on the known reflectance properties. In particular, the behaviour
of specular reflections subject to changing viewpoint and the resulting effects on the
estimation of disparities are taken into account. Furthermore, it is possible to utilise
the method for large baseline stereo, where the stereo baseline is comparable to the
object distance, as the differences in perspective introduced by the strongly differ-
ent viewpoints are explicitly considered as well. All available triangulation-based
and photopolarimetric information is utilised for the purpose of three-dimensional
surface reconstruction.
The iterative optimisation scheme described in this section assumes convergence.
However, the solution is not guaranteed to converge, since the triangulation-based
and photopolarimetric information may in principle be contradictory, e.g. due to
camera calibration errors or inaccurate knowledge of the illumination direction.
However, we observed that in all experiments regarded in Sect. 6.3.4 convergence is
achieved after 4-8 iteration cycles. Self-consistent measures for assessing the three-
dimensional reconstruction accuracy and convergence behaviour are discussed in
Sect. 6.3.4.3 .
5.4.2 Qualitative Behaviour of the Specular Stereo Algorithm
Figure 5.20 shows the observed vs. the rendered images of the specular surface of
the connection rod. While initially in camera 1 the rendered image is very similar to
the observed image since the initial SfPRD step aims for obtaining a small value of
e I according to ( 3.19 ), the differences between the observed image in camera 2 and
the correspondingly rendered image are evident. During the initial SfPRD step, the
intensity information of image 2 is not taken into account. The differences in surface
shape occur because the large-scale surface slope in the vertical image direction is
inaccurately estimated by the initial SfPRD step—for only less than 0 . 5%ofthe
image pixels can an initial disparity value be determined. Differences in surface
brightness (the right third of the surface profile appears bright in the observed image
but fairly dark in the rendered image) are due to the same large-scale inaccuracies
of the initial three-dimensional shape estimate, which result in inaccurate surface
gradients and thus estimated surface brightness.
After 8 cycles of the iterative scheme proposed in Sect. 5.4.1 , the geometric ap-
pearance of the surface and the distribution of surface brightness in both rendered
images closely resemble the corresponding observed images. Now a disparity value
can be determined for 18 % of the image pixels. This example qualitatively illus-
trates the convergence behaviour of the proposed algorithm. Section 6.3.4 provides