Efficient stereo with multiple windowing

Abstract

We

Adaptive Scale Selection for Hierarchical Stereo

“…However, the limitations of block-based matching are exacerbated while operating in the discrete scale space, leading to both heavy distortions near surface boundaries as well as loss of fine detail structures. The work of Sizintsev [12] imple- ments a hierarchical stereo approach, which incorporates adaptive matching window support across scales, in a manner similar to shiftable windows [1]. While such an approach improves border localization, it does not provide a mechanism for fine structure recovery.…”

Dense disparity maps from sparse disparity measurements

“…These methods assume constant disparity over the support, which does not hold at disparity discontinuities and leads to boundary fattening artifacts. Explicit occlusion detection [11], multiple windowing [14], or adaptive local support weighting [29] reduce this effect, but cannot avoid it completely.…”

Efficient Stereo with Accurate 3-D Boundaries

“…The latter step is similar to morphological operation on the match score map (dilation for Normalized Cross Correlation match measure) using the aggregation window as a structural element to simulate shiftable windows in single-scale matching [14]. Note that the proposed approach is not identical to estimating disparity estimates at each level via shiftable windows [8], because, for each pixel, each shifted window corresponds to a different disparity offset.…”

“…While the first flaw of block matcher is well-researched and a number of remedies exist via introduction of shiftable/overlapping/adaptive windows [8,9,19,21], the second one is specific to CTF and has not been paid enough attention. Thus, in the following, more discussion is devoted to the disparity upsampling procedure.…”

“…In the past, much research has considered recovery of binocular disparity near 3D boundaries [14,5,1]. For local methods, the use of adaptive spatial support for match windows can ameliorate issues arising in attempts to match near 3D discontinuities by shaping windows to avoid poorly defined matches [8,9,19,21]. Many recent advances in disparity estimation near 3D boundaries explicitly consider half-occlusions, where one view sees portions of a background surface that are occluded to the other view by a foreground surface.…”