Two approaches to computing view-dependent depth maps are presented in this research paper. In the first approach, a depth map is computed for each image, whereas in the second approach, depth maps are computed simultaneously for frames chosen with various criteria.
Common to both approaches is the computation of matching costs refined with sum of sum of squared differences (SSSDs). Additionally, shiftable windows are used in order to alleviate such problems as occlusions. The authors rightly point out that a temporal shift is equivalent to a spatial one, in the case of partial occlusions.
The authors explore global techniques, and define an energy function to be optimized. They combine occluded pixel labeling, visibility reasoning, and hierarchical disparity computation using graph cuts to obtain depth estimates.
The numerical results are convincing: the two techniques work well, in particular for some challenging sequences such as the Garden Flower, which contains significant occlusions. However, the paper is hard to approach, mainly because some of the writing and figure captions are unclear.