Dense 3D Reconstruction and Tracking of Dynamic Surface - Advances in Image and Graphics Technologies - page 221

Fig. 6. Reconstruction and tracking of dynamic cloth
the correct solution can be easily achieved, but for those with sparse texture, large
block size is needed in order to gain the correct iterative solution.
Due to their different physical properties, the behavior of the surfaces ranges
from smooth deformations for the fish surface to sharp folds and creases for the cloth.
However, no parameter tuning was needed to obtain these results with our
algorithm. Fig.4 shows the reconstructing and the tracking of dynamic fish surface,
(a) is the stereo images in which the top two sub-images are the images of previous
frame in left and right image sequence, and the bottom two are those of the next
frame, (b) is the reconstructed 3D surface, and (c) is the motion field calculated by the
tracking information between the previous and next 3D surface which includes the
accurate orientation and distance of the motion of every point in the previous 3D
surface. Therefore, this method provides a useful approach to those applications such
as analysis of motion rules of the dynamic surface. Fig.5 and Fig.6 a are the results of
paper and cloth respectively.
The deformation of dynamic fish surface and paper is usually smoothing as
shown in Fig.4 and Fig.5, which deforms smoothly when they are moving, so it is
more precise to capture the dynamic data for the motion process. The dynamic
process of cloth motion is usually very sharp; in this case, there may exist bigger error
in reconstruction and tracking. In order to solve this problem, the neighboring blocks
can be overlapped with each other, as mentioned in the previous section. Fig.6
demonstrates the reconstruction of the folded cloth, in which (a) shows the matching
without block overlapping, (b) and (c) are the corresponding reconstructed result and
the motion field, and (d) illustrates the matching with 50% block overlapping, (e) and
(f) are the result and motion field. By comparing (b) and (e), we realized that the
overlapping case produces better results.