Accurate dense stereo by constraining local consistency on superpixels (ICPR 2010)

Segmentation is a low-level vision cue often deployed by stereo algorithms to assume that disparity within superpixels varies smoothly. In this paper, we show that constraining, on a superpixel basis, the cues provided by a recently proposed technique [1], which explicitly models local consistency among neighboring points, yields accurate and dense disparity fields. Our proposal, starting from the initial disparity hypotheses of a fast dense stereo algorithm based on scanline optimization, demonstrates its effectiveness by enabling us to obtain results comparable to top-ranked algorithms based on iterative disparity optimization methods.

S. Mattoccia, "Accurate dense stereo by constraining local consistency on superpixels"
, 20th International Conference on Pattern Recognition (ICPR2010), August 23-26, 2010, Istanbul, Turkey [Pdf] [Supplementay_material] [Bibtex] 


Under construction...

Experimental results according to the Middlebury automatic evaluation [2,3] deploying the initial disparity hypotheses provided by algorithm [4], available in [3] and  referred to as DS. The output of the proposed algorithm is referred to as D(3). Further details can be found in the paper and in the slides linked at the bottom of this page.

Evaluation on the Middlebury dataset

Results according to the Middlebury evaluation

Experimental results on the Venus stereo pair

Complete experimental results on the Venus stereo pair [2,3] using the initial disparity hypothesses provided by [4] and available on the same website. For segmentation I deployed the Mean-Shift algorithm [5] (source code available in [6]).

Reference      Target
Reference                                                                       Target

Initial disparity Hypotheses provided by SGM      Initial disparity hypotheses processed by the LC technique
           Initial disparity hypotheses DS (SGM [4])                                         After LC [1]                              

Oversegmentation (Mean-Shift)      Phase 1: constraining local consistency on (oversegmented) superpixels
Oversegmentation [5,6]                                                          Phase 1             

Segmentation (Mean-Shift)      Phase 2: constraining local consistency on (segmented) superpixels
Relaxing segmentation [5,6]                                                    Phase 2                

Final disparity map
Proposed method

If you are interested in stereo vision, you might find interesting this:

    Stefano Mattoccia
    "Stereo vision: algorithms and applications"
    Extended version of the talk given at the University of Twente, April 1st 2009

If you have any question feel free to contact me at:

Stefano Mattoccia, email: smatt AT ieee DOT org


[1] S. Mattoccia, "A locally global approach to stereo correspondence", IEEE Workshop on 3D Digital Imaging and Modeling  (3DIM2009), October 3-4, 2009, Kyoto, Japan [Abstract] [Pdf] [Bibtex]

[2] D. Scharstein and R. Szeliski, "A taxonomy and evaluation of dense two-frame stereo correspondence algorithms",International Journal of Computer Vision, 47(1/2/3):7-42, April-June 2002 / Microsoft Research Technical Report MSR-TR-2001-81, November 2001

[3] D. Scharstein and R. Szeliski, Middlebury Stereo Vision Page, URL: ""

[4] H. Hirschmüller, "Stereo vision in structured environments by consistent semi-global matching", CVPR 2006, PAMI 30(2):328-341, 2008

[5] D. Comaniciu and P. Meer,  "Mean shift: a robust approach toward feature space analysis", PAMI, 24(5), 603-619,2002

[6] C. Christoudias, B. Georgescu, P. Meer,  EDISON:


last update on: Aug 29, 2010