3.5  Summary and conclusions

Using the geometry of multiple views, we have derived the basic principles employed to calculate the depth using two views and multiple views. We have discussed two popular optimization strategies employed for estimating disparity images: WTA and and a one-dimensional optimization strategy based on dynamic programming. In the presented experiments, we have highlighted three important disadvantages of the two previously mentioned depth-estimation algorithms: (1) only two rectified views are employed simultaneously, (2) consistency of depth estimates across the views is not enforced and (3) multiple pre-processing (image rectification) and post-processing (disparity-to-depth conversion and image de-rectification) procedures are necessary. To address these problems, in the last section, we have introduced an alternative depth image estimation algorithm. The proposed algorithm is based on two novel constraints, i.e., an inter-line and an inter-view smoothness constraint, which enforce smooth variations of depth values across scanlines and consistent depth values across the views. The first pass of the algorithm serves as an estimate for depth image initialization, whereas the second pass refines the initial depth images by enforcing consistent depth across the views. Experiments have shown that especially the second pass provides a noticeable quality improvement by detecting and removing fuzzy outliers that are not consistent across the views. We have shown than both smoothness constraints can be efficiently integrated into the one-dimensional optimization dynamic programming algorithm. Experiments have shown that the proposed constraints yield reasonably accurate depth values for the critical case of texture-less and colorless multi-view images.

At this point, it is relevant to come back on the initial requirements stated in the introduction of this chapter. The primary requirement of the depth estimation algorithm is that it should lead to accurate depth images featuring smooth regions delineated by sharp edges (piecewise linear regions). To fulfill this requirement, we have proposed a post-processing step that segments the texture image into regions of linear depth. The above-discussed piecewise linear property of the depth images is highly relevant for the coding of depth images, as discussed in Chapter 6. The second important contribution and initial requirement is that the resulting depth images should be consistent across the views. The proposed two-pass algorithm explicitly detects and removes inconsistent depth estimates to enforce consistency. It will be seen in Chapter 5, that such an inter-view consistency can be beneficially exploited for compression purposes, where the inter-view consistency corresponds to an inter-view redundancy.