Improving the accuracy of fast dense stereo correspondence algorithms by enforcing local consistency of disparity fields

Accurate, dense 3D reconstruction is an important requirement in many applications, and stereo represents a viable alternative to active sensors. However, top-ranked stereo algorithms rely on iterative 2D disparity optimization methods for energy minimization that are not well suited to the fast and/or hardware implementation often required in practice. An exception is represented by the approaches that perform disparity optimization in one dimension (1D) by means of scanline optimization (SO) or dynamic programming (DP). Recent SO/DP-based approaches aim to avoid the well known streaking effect by enforcing vertical consistency between scanlines deploying aggregated costs, aggregating multiple scanlines, or performing energy minimization on a tree. In this paper we show that the accuracy of two fast SO/DP-based approaches can be dramatically improved by exploiting a non-iterative methodology that, by modeling the coherence within neighboring points, enforces the local consistency of disparity fields. Our proposal allows us to obtain top-ranked results on the standard Middlebury dataset and, thanks to its computational structure and its reduced memory requirements, is potentially suited to fast and/or hardware implementations.