1 / 38

A Local Adaptive Approach for Dens e Stereo Matching i n Architectural Scene Reconstruction

A Local Adaptive Approach for Dens e Stereo Matching i n Architectural Scene Reconstruction. C . Stentoumis 1 , L. Grammatikopoulos 2 , I. Kalisperakis 2 , E. Petsa 2 , G. Karras 1 1. Laboratory of Photogrammetry, Department of Surveying,

jacoba
Download Presentation

A Local Adaptive Approach for Dens e Stereo Matching i n Architectural Scene Reconstruction

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. A Local Adaptive Approach for Dense Stereo Matching in Architectural Scene Reconstruction C. Stentoumis1, L. Grammatikopoulos2, I. Kalisperakis2, E. Petsa2, G. Karras1 1. Laboratory of Photogrammetry, Department of Surveying, National Technical University of Athens, GR-15780 Athens, Greece 2. Laboratory of Photogrammetry, Department of Surveying, Technological Educational Institute of Athens, GR-12210 Athens, Greece 5th International Workshop on 3D Virtual Reconstruction and Visualization of Complex Architectures (3D-ARCH '2013), 25-26 February 2013

  2. Outline • Introduction • Related Work • Proposed Algorithm • Experimental Results • Conclusion

  3. Introduction

  4. Objective • Present a method • Combines pre-existing algorithms and novel considerations • With good sub-pixel accuracy

  5. Related work

  6. Related Work Stereo Matching : 5 14 (X,Y) (X-d,Y) Range: 0 - 16

  7. Related Work

  8. ProposedAlgorithm

  9. Census on intensity principal derivatives • Census transformation based on gradients: • Less sensitive to radiometric differences and repetitive patterns Intensity Gradient

  10. Census on Gradients Census transform window :

  11. Hamming Distance • Left image • Right image Hamming Distance = 3 XOR

  12. Comparison 2.5% less erroneous pixels • After aggregation step: Default census Census on gradients

  13. Comparison • After aggregation step[13]: [13]Mei X., Sun X., Zhou M., Jiao S., Wang H., Zhang X., 2011. On building an accurate stereo matching system on graphics hardware. ICCV Workshop on GPU in Computer Vision Applications.

  14. Absolute Difference on Image Color and Gradients AD ( color ) : AD ( Gradient ) :

  15. Total Matching Cost • normalized by λ Census (gradient) AD (color) AD (gradient)

  16. Census (gradient) AD (color) AD (gradient) Combined

  17. [25]Zhang K., Lu J., Lafruit G., 2009. Cross-based local stereo matching using orthogonal integral images. IEEE Transactions on Circuits & Systems for Video Technology. Support Region • Cross-based support region[25]: • Threshold of cross-skeleton expansion:

  18. [16]Stentoumis C., Grammatikopoulos L., Kalisperakis I., Karras G., 2012. Implementing an adaptive approach for dense stereo matching. International Archives of Photogrammetry, Remote Sensing & Spatial Information Sciences, Support Region • Threshold of cross-skeleton expansion[16]: • Lmax: largest semi-dimension of the window size • τmax : largest color dissimilarity between p and q 3X3 median filter P q lq

  19. Support Region

  20. Aggregation Step • A. Normalized by the number of pixels in the support region • B. 3D Gaussian function is applied for smoothing the aggregated costs. • C. Winner-take-all

  21. Comparison • After aggregation step[13]: • Run this step for 4 iterations to get stablecost values. • For iteration 1 and 3, aggregated horizontally and thenvertically. • For iteration 2 and 4, aggregated verticallyand then horizontally.

  22. [13] method Proposed Method

  23. Refinement • Left-right consistency check • Pixel p is characterized as valid (inlier) if the following constraint holds:

  24. Refinement • Outlier cross-based filtering • The cross-based support regions provide a robust description of pixel neighborhoods • The median value of inliers in the support region is selected and attributed to the mismatched pixel.

  25. Refinement • Occlusion / mismatch labeling • Remaining outliers are re-estimated • Mismatches: • The epipolar line of the mismatch pixel intersects with disparity function • Use median interpolation in a small patch around them • Occlusions • Use the second lowest disparity value in the neighborhood

  26. Refinement • EpipolarLine Before After

  27. Refinement • Sub-pixel estimation • Estimation at the sub-pixel level is made by interpolating a 2nd order curve to the cost volume C(d). • This curve is defined by the disparitiesof the preceding and following pixelsand their corresponding cost values • Choose minimum cost position through a closed form solution for the 3 curve points.

  28. Refinement • Disparity map smoothing • Median filter is applied. The effect of overall post-processing refinement

  29. ExperimentalResults

  30. Experimental Results • Evaluated on the Middlebury and EPFL multi-view datasets • Parameter values were kept constant for all tests.

  31. Experimental Results • Middlebury evaluation Error Threshold = 1 Error Threshold = 0.75

  32. [13] method Proposed Method

  33. Experimental Results Threshold = 0.75 , % of wrong pixels

  34. Experimental Results Herz-Jesu-K7 stereo pair

  35. Experimental Results Herz-Jesu-K7 stereo pair

  36. Conclusion

  37. Conclusion

More Related