Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 80

a d = 0 . 006 and d max = 1 . 2 pixels for the outdoor scene and a d = 0 . 060 and

d max = 12 pixels for the indoor scenes.

Based on this error analysis, in Sect. 1.6.4 the minimum wavelength of the repet-
itive structures allowed to apply the proposed method is estimated for the outdoor
and the indoor scenario, respectively.

1.6.4 Experimental Evaluation

In this section we describe an experimental evaluation of the proposed method for
resolving stereo matching errors. For image acquisition, we utilised a Point Grey
Digiclops camera system with an image size of 1024 × 768 pixels, a camera constant

of 6 mm (corresponding to f 100 mm.

The images were rectified to standard epipolar geometry based on the algorithm
by Fusiello et al. ( 2000 ). We regard three different scenes, each showing a small
object in front of a pronounced repetitive structure. The right image of each stereo
pair is shown in Fig. 1.26 a. Scene 1 displays a person standing in front of a large
fence, scene 2 shows a keyboard with a hand in front of it, scene 3 displays an arm
showing repetitive structures due to a pattern on the clothes with a bar in front of
it, and scene 4 shows a typical urban environment with a person standing in front
of a building. Scenes 1 and 4 are outdoor scenes while scenes 2 and 3 are indoor
scenes. For scenes 1, 2, and 4 the plane model (cf. Sect. 1.6.1 ) is used; for scene 3
the articulated hand-arm model (cf. Sect. 1.6.2 ) is used.
For quantitative evaluation of the resulting three-dimensional point clouds, we
have generated ground truth data by manual labelling (cf. Fig. 1.26 ). For each initial
three-dimensional point cloud, the adaptation result of the plane model (scenes 1, 2,
and 4) and the hand-arm model (scene 3) is shown in Fig. 1.27 . In all four examples,
the model adaptation is performed accurately, and the spurious objects arising due to
the repetitive structures are clearly evident. The fact that the local stereo algorithm
is not able to provide an appropriate three-dimensional reconstruction of the scene
would lead to an unsatisfactory behaviour of subsequent processing stages such as
scene segmentation and object detection, e.g. in a mobile robotic system, as the
spurious objects would then result in spurious obstacles.
For comparison, the result of the model-based stereo approach is shown in
Fig. 1.28 for scene 1. This technique yields a good three-dimensional reconstruc-
tion of the fence but of course fails to capture the person standing in front of it, as
it is not part of the model. However, an initial pose which is already fairly close
to the final result has to be known a priori, where the initial configuration shown
in Fig. 1.28 already corresponds to the maximum offset for which the model-based
stereo algorithm converges. In contrast, the plane model adaptation approach de-
scribed in Sect. 1.6.1 does not require initial values for the pose parameters. Hence,
both complementary standard approaches (local and model-based stereo) cannot
provide a satisfactory simultaneous three-dimensional reconstruction of the scene
part displaying repetitive structures and the object in front of it.

= 1350 pixels), and a baseline distance of l =