Three-Dimensional Pose Estimation and Segmentation Methods - 3D Computer Vision: Efficient Methods and Applications - page 106

Fig. 2.4 ( a ) One of the three artificially rendered input images with the overlaid reprojec-
tions of the three-dimensional ribbon snake. A virtual trinocular camera with a resolution of
1024 768 pixels and a base distance of 100 mm has been used to generate the images. ( b )Re-

constructed tube, shown from a different viewpoint. The RMSE between ground truth and recon-
struction is 1.5 mm

×

Fig. 2.5 Behaviour of the three-dimensional ziplock ribbon snake algorithm with respect to partial
occlusion of the object. A virtual rod is moved across the rendered scene of Fig. 2.4

centre curve of the estimated ribbon snake with respect to the model centre curve.
In the example shown in Fig. 2.4 , the RMSE amounts to 1 . 5 mm, which roughly
corresponds to 1 pixel disparity error. Subpixel greedy stepwidths and interpolated
image energy calculation were used to obtain this result.
The behaviour of the algorithm in the presence of partial occlusions has been
tested by moving a virtual rod over the scene, as shown in Fig. 2.5 .InFig. 2.5 a,
where only a small part of the object is occluded, the RMSE with respect to the
ground truth amounts to 1 . 1 mm, while for stronger occlusions as in Figs. 2.5 b
and c the RMSE corresponds to 3 . 1 mm and 3 . 8 mm, respectively. All described test
cases were run on a 1 . 7 GHz Pentium Mobile Processor. The computation time of
the optimisation procedure amounts to between 1 and 23 seconds, depending on the
complexity of the scene and the parameters chosen for the optimisation procedure.
The experimental results show that the proposed three-dimensional ziplock rib-
bon snake algorithm is able to perform a fairly accurate three-dimensional contour
segmentation. The stability of the algorithm is mainly due to the usage of model-
based constraints and initial contour curves based on model information. However,