Integrated Frameworks for Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 194

Fig. 5.6 Three-dimensional reconstruction of the lava stone

5.1.3.2 Evaluation of the Online Algorithm

The online algorithm is evaluated by Wöhler et al. ( 2009 ), where the same image
sequences are used as for the offline algorithm. Initially, the three-dimensional point
cloud reconstructed by the online algorithm is very noisy, as only a small number of
features have already reached their best focus. The three-dimensional reconstruction
result then becomes increasingly similar to the result of the offline algorithm, where
small differences remain due to the fact that slightly different indices f i according
to ( 4.16 ) are determined for the best focused ROI of each tracked feature. For all
image sequences, the same reference distances as for the evaluation of the offline
algorithm were used (distances between reference points for the cuboid sequence
and for the lava stone sequence, bottle diameter for the bottle sequence). The error
bars in Figs. 5.7 , 5.9 , and 5.10 denote the standard deviations over 100 online runs,
respectively, where Gaussian noise was added to the images before each online run.
For the cuboid sequence (cf. Fig. 5.8 ), the first three-dimensional reconstruction
result is obtained when 21 images are processed. In Fig. 5.7 , the estimated distance
between pairs of reference points is shown as soon as they have passed their re-
spective point of best focus. The correspondingly estimated distances are used to
determine the average relative scale error shown in Fig. 5.7 . An intuitive finding in-
dicated by Fig. 5.8 is that the three-dimensional reconstruction result becomes more
and more accurate with increasing number of processed images and tracked features
having passed their point of best focus.
Similar evaluations were performed for the bottle sequence and for the lava stone
sequence (cf. Fig. 5.9 ). The estimated absolute scale of the bottle deviates by less
than 3 . 0 % from the true scale once 12 images have been processed, then decreases
further and increases slightly again, while the standard deviation decreases con-
stantly. When all images of the sequence are processed, the error of the estimated
absolute scale corresponds to 1 . 9 standard deviations.
For the lava stone sequence, the absolute scale is reconstructed at a relative accu-
racy of 1 . 2 %, corresponding to one standard deviation, after 12 processed images.
However, when the image sequence is processed completely, the deviation between