Volume III-3
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., III-3, 123-130, 2016
https://doi.org/10.5194/isprs-annals-III-3-123-2016
© Author(s) 2016. This work is distributed under
the Creative Commons Attribution 3.0 License.
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., III-3, 123-130, 2016
https://doi.org/10.5194/isprs-annals-III-3-123-2016
© Author(s) 2016. This work is distributed under
the Creative Commons Attribution 3.0 License.

  03 Jun 2016

03 Jun 2016

REVISITING INTRINSIC CURVES FOR EFFICIENT DENSE STEREO MATCHING

M. Shahbazi1, G. Sohn2, J. Théau1, and P. Ménard3 M. Shahbazi et al.
  • 1Dept. of Applied Geomatics, Université de Sherbrooke, Boul. de l'Université, Sherbrooke, Québec, Canada
  • 2Dept. of Geomatics Engineering, York University, Keele Street, Toronto, Ontario, Canada
  • 3Centre de géomatique du Québec, Saguenay, Québec, Canada

Keywords: Matching, Intrinsic Curves, Search Space Reduction, Occlusion, Energy, Optimization

Abstract. Dense stereo matching is one of the fundamental and active areas of photogrammetry. The increasing image resolution of digital cameras as well as the growing interest in unconventional imaging, e.g. unmanned aerial imagery, has exposed stereo image pairs to serious occlusion, noise and matching ambiguity. This has also resulted in an increase in the range of disparity values that should be considered for matching. Therefore, conventional methods of dense matching need to be revised to achieve higher levels of efficiency and accuracy. In this paper, we present an algorithm that uses the concepts of intrinsic curves to propose sparse disparity hypotheses for each pixel. Then, the hypotheses are propagated to adjoining pixels by label-set enlargement based on the proximity in the space of intrinsic curves. The same concepts are applied to model occlusions explicitly via a regularization term in the energy function. Finally, a global optimization stage is performed using belief-propagation to assign one of the disparity hypotheses to each pixel. By searching only through a small fraction of the whole disparity search space and handling occlusions and ambiguities, the proposed framework could achieve high levels of accuracy and efficiency.