Integrated Frameworks for Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 193

Ta b l e 5 . 1 Summary of the evaluation results for the offline version of the combined structure from
motion and depth from defocus algorithm. The indicated error intervals correspond to the standard
deviations
Sequence

Reference length (mm)
Ground truth

Length
(images)

E reprojection
(pixels)

E defocus
(pixels)

Reconstruction

Cuboid 46 0 . 642 0 . 636 32 . 0 34 . 1 ± 1 . 6

Bottle 26 0 . 747 0 . 387 80 . 0 82 . 8 ± 1 . 4

Lava stone 15 0 . 357 0 . 174 60 . 0 58 . 3 ± 0 . 8

Fig. 5.5 Three-dimensional reconstruction of the lower, cylindrical part of the bottle

background features are selected, since none of these features obtains its maximum
sharpness in the acquired sequence. The three-dimensional reconstruction result is
shown in Fig. 5.5 . The resulting reprojection error corresponds to E reprojection =
0 . 75 pixel and the defocus error to E defocus = 0 . 39 pixel. For analysing the absolute

scale, the reconstructed three-dimensional point cloud was projected on the ground
plane, and a circle was adapted to the projected points. The resulting bottle diameter
corresponds to 82 . 8 mm, which is consistent with the true diameter of 80 . 0mm(cf.
Fig. 5.5 and Table 5.1 ).

Lava Stone Sequence As a further real-world object, the lava stone shown in
Fig. 5.2 c is examined. The three-dimensional reconstruction result is shown in
Fig. 5.6 . A Delaunay triangulation of the three-dimensional point cloud was used
to generate the shaded view of the object. The cusp visible in the left part of
the reconstructed surface is the result of three outlier points generated by inaccu-
rately determined feature positions. The resulting reprojection error corresponds to
E reprojection =

0 . 357 pixel and the defocus error to E defocus = 0 . 174 pixel. As a ref-

erence for the absolute scale, a pair of well-defined points with a true distance of
60 . 0 mm was chosen. The estimated distance corresponds to 58 . 3 mm, which is in
reasonable correspondence with the true value (cf. Table 5.1 ).