Iterative Semi-Global Matching [iSGM]

Iterative Semi-Global Matching [iSGM] Submitted on 27 Jun. 2012 05:56 by Simon Hermann (The University of Auckland) Running time:   8 s Environment:   2 cores @ 2.5 Ghz (C/C++) Method Description: The conceptual parallel scan-line integration strategy of standard SGM is translated into an iterative scheme. The accumulation buffer is evaluated to a prior disparity map after horizontal and after vertical integration. The results serve as input to generate a novel data structure, the semi-global distance map (SGDM), that allows a fast analysis of spatial disparity information. This information is used to reduce the search-space and also increases the robustness of the final solution, especially in case of weak underlying stereo data. The method is coupled with a coarse-to-fine approach where  costs rather than disparities are upscaled. Full left-right  consistency checks are performed throughout the method. Run-time without SSE-optimization and without GPU usage. Parameters: P1=35 P2=350 search space=[0,255] Latex Bibtex: @INPROCEEDINGS{HERM2012,  author = {Simon Hermann and Reinhard Klette},  title = {Iterative Semi-Global Matching for Robust Driver  Assistance Systems},  booktitle = {ACCV},  year = {2012}  }

Detailed Results

This page provides detailed results for the method(s) selected. For each of the first 20 test images, the number of erroneous pixels at all thresholds is depicted in the table. Underneath, the left input image, the disparity / end-point error map and the estimated (and interpolated) disparity / optical flow map are shown. The error map scales linearly between 0 (black) and >=5 (white) pixels error. Red denotes all occluded pixels, falling outside the image boundaries. The false color map is scaled to the largest ground truth disparity / flow value.

Test Set Average

Error Out-Noc Out-All Avg-Noc Avg-All 2 pixels 7.94 % 10.00 % 1.2 px 2.1 px 3 pixels 5.11 % 7.15 % 1.2 px 2.1 px 4 pixels 3.84 % 5.82 % 1.2 px 2.1 px 5 pixels 3.13 % 5.02 % 1.2 px 2.1 px This table as LaTeX

Reflective Regions

Error Out-Noc Out-All Avg-Noc Avg-All 2 pixels 31.87 % 35.72 % 4.7 px 8.6 px 3 pixels 24.67 % 28.86 % 4.7 px 8.6 px 4 pixels 20.52 % 24.84 % 4.7 px 8.6 px 5 pixels 17.73 % 22.10 % 4.7 px 8.6 px This table as LaTeX

Test Image 0

Error Out-Noc Out-All Avg-Noc Avg-All 2 pixels 7.87 % 10.52 % 1.0 px 1.9 px 3 pixels 5.59 % 8.22 % 1.0 px 1.9 px 4 pixels 4.36 % 6.89 % 1.0 px 1.9 px 5 pixels 3.46 % 5.84 % 1.0 px 1.9 px This table as LaTeX

Test Image 1

Error Out-Noc Out-All Avg-Noc Avg-All 2 pixels 6.53 % 10.21 % 0.9 px 2.2 px 3 pixels 3.93 % 7.57 % 0.9 px 2.2 px 4 pixels 3.00 % 6.38 % 0.9 px 2.2 px