The pairwise disparity estimation allows to compute image to image correspondences between adjacent rectified image pairs, and independent depth estimates for each camera viewpoint. An optimal joint estimate will be achieved by fusing all independent estimates into a common 3D model. The fusion can be performed in an economical way through controlled correspondence linking as described in this section. The approach utilizes a flexible multi-viewpoint scheme by combining the advantages of small baseline and wide baseline stereo.
As small baseline stereo we define viewpoints where the baseline is much smaller than the observed average scene depth. This configuration is usually valid for image sequences were the images are taken as a spatial sequence from many slightly varying view-points. The advantages (+) and disadvantages (-) are
+ easy correspondence estimation, since the views are similar,
+ small regions of viewpoint related occlusionsG1,
- small triangulation angle, hence large depth uncertainty.
The wide baseline stereo in contrast is used mostly with still image photographs of a scene where few images are taken from a very different viewpoint. Here the depth resolution is superior but correspondence and occlusion problems appear:
- hard correspondence estimation, since the views are not similar,
- large regions of viewpoint related occlusions,
+ big triangulation angle, hence high depth accuracy.
The multi-viewpoint linking combines the virtues of both approaches. In addition it will produce denser depth maps than either of the other techniques, and allows additional features for depth and texture fusion. Advantages are:
+ very dense depth maps for each viewpoint,
+ no viewpoint dependent occlusions,
+ highest depth resolution through viewpoint fusion,
+ texture enhancement (mean texture, highlight removal, super-resolution texture).