Efficient Large-Scale Stereo Matching
Andreas Geiger
1
, Martin Roser
1
, and Raquel Urtasun
2
1
Dep. of Measurement and Control, Karlsruhe Institute of Technology
2
Toyota Technological Institute at Chicago
geiger@kit.edu, martin.roser@kit.edu, rurtasun@ttic.edu
Abstract. In this paper we propose a novel approach to binocular stereo
for fast matching of high-resolution images. Our approach builds a prior
on the disparities by forming a triangulation on a set of support points
which can be robustly matched, reducing the matching ambiguities of
the remaining points. This allows for efficient exploitation of the dispar-
ity search space, yielding accurate dense reconstruction without the need
for global optimization. Moreover, our method automatically determines
the disparity range and can be easily parallelized. We demonstrate the
effectiveness of our approach on the large-scale Middlebury benchmark,
and show that state-of-the-art performance can be achieved with signif-
icant speedups. Computing the left and right disparity maps for a one
Megapixel image pair takes about one second on a single CPU core.
1 Introduction
Estimating depth from binocular imagery is a core subject in low-level vision as
it is an important building block in many domains such as multi-view reconstruc-
tion. In order to be of practical use for applications such as autonomous driving,
disparity estimation methods should run at speeds similar to other low-level vi-
sual processing techniques, e.g. edge extraction or interest point detection. Since
depth errors increase quadratically with the distance [1], high-resolution images
are needed to obtain accurate 3D representations. While the benefits of high
resolution imagery are already exploited exhaustively in structure-from-motion,
object recognition and scene classification, only few binocular stereo methods
deal efficiently with large images.
Stereo algorithms based on local correspondences [2, 3] are typically fast,
but require an adequate choice of window size. As illustrated in Fig. 1 this
leads to a trade-off between low matching ratios for small window sizes and
border bleeding artifacts for larger ones. As a consequence, poorly-textured and
ambiguous surfaces cannot be matched consistently.
Algorithms based on global correspondences [4–9] overcome some of the afore-
mentioned problems by imposing smoothness constraints on the disparities in the
form of regularized energy functions. Since optimizing such MRF-based energy
functions is in general NP-hard, a variety of approximation algorithms have been
proposed, e.g., graph cuts [4, 5] or belief propagation [6]. However, even on low-
resolution imagery, they generally require large computational efforts and high
