Triangulation-Based Approaches to Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 84

Ta b l e 1 . 3 Number of extracted 3D points for the regarded example scenes when neglecting model
information ( λ e = 0) and taking it into account ( λ e = 0 . 01)
Scene Constraint λ e = 0 λ e = 0 . 01

Fence with person

Uniqueness 23317 24755

Uniqueness + ordering 15509 22752

Min. weighted matching 25354 25463

Keyboard with hand

Uniqueness 9392 9924

Uniqueness + ordering 7613 9029

Min. weighted matching 9980 10145

Arm with bar

Uniqueness 9164 9587

Uniqueness + ordering 8329 8601

Min. weighted matching 9604 10064

Building with person

Uniqueness 5023 5651

+

Uniqueness ordering 3270 5136

Min. weighted matching 5648 5654

between 0 . 001 and 0 . 01 for all four example scenes. Hence, the choice of the pa-
rameter λ e does not critically depend on the regarded scene.
In contrast to classical outlier rejection approaches, applying the proposed tech-
nique does not decrease the absolute number of three-dimensional points but actu-
ally increases it for all four regarded scenes. This increase is most significant when
using the combined uniqueness and ordering constraint (cf. Table 1.3 ).
In Fig. 1.30 , the resulting three-dimensional point clouds are shown in disparity
space for the four example scenes, setting λ e = 0 (i.e. model information is ne-

glected) and λ e = 0 . 01. In scene 1 (fence with person, cf. Fig. 1.30 a), only a small

number of 'spurious fence' points remain visible once we set λ e = 0 . 01, while the

object in front of the fence is extracted correctly. A similar behaviour occurs for
scene 2 (keyboard with hand, cf. Fig. 1.30 b), where an even smaller number of in-
correctly assigned points on the repetitive structures remains for λ e =

0 . 01, which
do not form clusters but are more or less evenly distributed in disparity space. In
scene 3 (arm and bar, cf. Fig. 1.30 c), the effect of the proposed method is less
pronounced than for scenes 1 and 2, but the density of the remaining incorrectly
assigned points on the repetitive structure still decreases, and especially the clus-
ters of three-dimensional points clearly apparent for λ e = 750 pixels and

disparities of 140 and 175 pixels, respectively, disappear or become less dense for
λ e =

0at u ≈

0 . 01. The number of incorrectly assigned points on the object increases (these
points are assigned a disparity that approximately corresponds to that of the arm),
but the object between the camera and the arm remains represented by two dense
clusters of three-dimensional points. For scene 4 (urban scene showing a person in
front of a building, cf. Fig. 1.30 d), increasing λ e to values between 0 . 001 and 0 . 0025
strongly increases the fraction of correct points on the building, such that the spu-
rious objects disappear, while the fraction of correct points on the person decreases
only moderately.