Segmentation-Based Adaptive Support for Accurate Stereo Correspondence
Significant achievements have been attained in the field of dense stereo correspondence by local algorithms based on an adaptive support. Given the problem of matching two correspondent pixels within a local stereo process, the basic idea is to consider a
- PDF / 5,471,524 Bytes
- 12 Pages / 430 x 660 pts Page_size
- 20 Downloads / 214 Views
Department of Electronics Computer Science and Systems (DEIS) University of Bologna Viale Risorgimento 2, 40136 - Bologna, Italy 2 Advanced Research Center on Electronic Systems (ARCES) University of Bologna Via Toffano 2/2, 40135 - Bologna, Italy {ftombari, smattoccia, ldistefano}@deis.unibo.it
Abstract. Significant achievements have been attained in the field of dense stereo correspondence by local algorithms based on an adaptive support. Given the problem of matching two correspondent pixels within a local stereo process, the basic idea is to consider as support for each pixel only those points which lay on the same disparity plane, rather than those belonging to a fixed support. This paper proposes a novel support aggregation strategy which includes information obtained from a segmentation process. Experimental results on the Middlebury dataset demonstrate that our approach is effective in improving the state of the art. Keywords: Stereo vision, stereo matching, variable support, segmentation.
1 Introduction Given a pair of rectified stereo images Ir , It , the problem of stereo correspondence is to find for each pixel of the reference image Ir the correspondent pixel in the target image It . The correspondence for a pixel at coordinate (¯ x, y¯) can only be found at the same ¯ + dM ], where D = [dm , dM ] vertical coordinate y¯ and within the range [¯ x + dm , x denotes the so-called disparity range. The basic local approach selects, as the best correspondence for a pixel p on Ir , the pixel of It which yields the lowest score of a similarity measure computed on a (typically squared) fixed support (correlation window) centered on p and on each of the dM − dm candidates defined by the disparity range. The use of a spatial support compared to a pointwise score increases the robustness of the match especially in presence of noise and low-textured areas, but the use of a fixed support is prone to errors due to the fact that it blindly aggregates pixels belonging to different disparities. For this reason, incorrect matches tend to be generated along depth discontinuities. In order to improve this approach, many techniques have been proposed which try to select for each pixel an adaptive support which best aggregates only those neighbouring pixels at the same disparity [1], [2], [3], [4], [5], [6] (see [7] and [8] for a review). Recently very effective techniques [8], [9] were proposed, which represent state of the art D. Mery and L. Rueda (Eds.): PSIVT 2007, LNCS 4872, pp. 427–438, 2007. c Springer-Verlag Berlin Heidelberg 2007
428
F. Tombari, S. Mattoccia, and L. Di Stefano
for local stereo algorithms. The former technique weights each pixel of the correlation window on the basis of both its spatial distance and its colour distance in the CIELAB space from the central pixel. Though this technique provides in general excellent results, outperforming [9] on the Middlebury dataset1 , in presence of highly textured regions the support can shrink to a few pixels thus dramatically reducing the reliability of the
Data Loading...
