44
A Rapid Stereo Matching Algorithm Based on Disparity Interpolation Gang Yao Yong Liu Bangjun Lei Dong Ren Proceedings of2010 Conference on Dependable Compnting (CDC'2010) November 20-22, 2010, Yichang, China Institute of Intelligent Vision and Image Information China Three Gorges University

A Rapid S tereo M atching A lgorithm B ased on Disparity I nterpolation

  • Upload
    joanna

  • View
    29

  • Download
    0

Embed Size (px)

DESCRIPTION

A Rapid S tereo M atching A lgorithm B ased on Disparity I nterpolation. Gang Yao Yong Liu Bangjun Lei Dong Ren Proceedings of2010 Conference on Dependable Compnting (CDC'2010) November 20-22, 2010, Yichang , China Institute of Intelligent Vision and Image Information - PowerPoint PPT Presentation

Citation preview

Page 1: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

A Rapid Stereo Matching Algorithm Based on Disparity Interpolation

Gang YaoYong Liu

Bangjun LeiDong Ren

Proceedings of2010 Conference on Dependable Compnting (CDC'2010)November 20-22, 2010, Yichang, China

Institute of Intelligent Vision and Image Information

China Three Gorges University

Page 2: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Outline• Introduction•Method•Disparity Interpolation•Experiments•Conclusion

Page 3: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Introduction•The disparity and the matching cost are calculated

based on stereo matching▫left image as the base image.▫right image as the base image.

•The disparity in this two kind of situations are also optimized.

Page 4: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Introduction•Matching area and un-matching area can be further

marked.

•The disparity interpolation process is done ▫started from the edge of un-matching area.▫stop until all pixels in un-matching area are finished.

Page 5: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Introduction•The proposed algorithm in this paper on the basis of

the adaptive cost function in [13].

• Just to interpolation the results of local matching algorithm ▫maintain the disparity continuity.▫improve the stereo matching speed.

[13]Segment-Based Stereo Matching Using Belief Propagation and a Self-Adapting Dissimilarity Measure

Page 6: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Segment-Based Stereo Matching Using Belief Propagation and a Self-Adapting Dissimilarity Measure

Andreas KlausMario SormannKonrad Karner

VRVis Research Center, Graz, Austria

Page 7: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Introduction•Color segmentation and Self-adapting matching

▫maximize the number of reliable correspondences.

•Self-adapting conception : ▫The scene structure is modeled by a set of planar surface

patches.▫Instead of assigning a disparity value to each pixel, a

disparity plane is assigned to each segment.▫The optimal disparity plane labeling is approximated by

applying belief propagation.

Page 8: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Work-flow Diagram

Page 9: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Local Matching in Pixel Domain•.

• .

• .

Page 10: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Plane Estimation• .

▫step 1 : calculate matching cost.▫step 2 : the disparity plane with the minimum matching cost is assigned to each segment.▫step 3 : segments that are assigned to the disparity plane are grouped.

•Repeating the plane fitting for grouped regions.▫increase the accuracy

Page 11: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Plane Assignment•.

• .

• .

Page 12: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Experiments

Page 13: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

A Rapid Stereo Matching Algorithm Based on Disparity Interpolation

Back to …

Page 14: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

The Proposed Algorithm

Page 15: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Outline• Introduction•Method

▫Cost function▫Robust matching▫Disparity optimization

•Disparity Interpolation•Experiments•Conclusion

Page 16: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation
Page 17: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Cost Function•The cost function includes two parts :

▫C(x,y,d) = (1-ω) × Cdata(x,y,d) + ω × Csmooth(x,y,d)

•Data term▫.

•Smooth term▫.

Page 18: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Cost Function•Can ensure the gray consistency between the

correspondent pixels.

•Can guarantee the smooth consistency in the pixel neighbourhood▫find matching pixel accurately.▫improve the smoothness of disparity.

Page 19: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation
Page 20: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Robust Matching•The similarity cost function including

▫the gray consistency.▫the smoothness consistency in pixel neighbourhood.

•The cost function is actually a quadratic function▫like p(x) = x2, ▫its derivatives is p'(x) = 2x.

Page 21: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Robust matching•The noise affects is linear increase in stereo

matching.

•The noise is not control by the function, so the quantizing error and noise is very sensitive in both the data item and smooth item in cost function.

Page 22: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

• In order to reduce the affection of noise▫robustness matching function Lorentzian [16].

• the noise is control by Lorentzian function, and the influence is nearly zero

Robust Matching

Page 23: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Robust Matching•New cost function

▫.

• .

Page 24: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation
Page 25: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation
Page 26: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Optimization•No matter what kind of method, the corresponding

points can not be found for a part of the pixels in the image.

•Left occluded and right occluded cannot appear at the same time.

Page 27: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Optimization•The process of optimization as follows :

•Step 1 : ▫After robust matching method

right-to-left disparity dR-L(x,y) and left-to-right disparity dL-R(x,y).

value of cost function is calculated respectively as CR-

L(x,y) and CL-R(x,y).

Page 28: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Optimization•Step 2 :

▫dR-L(x,y)、 dL-R(x,y)、 CR-L(x,y)、 CL-R will be optimized.

occluded

Page 29: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation
Page 30: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Outline• Introduction•Method•Disparity Interpolation•Experiments•Conclusion

Page 31: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation
Page 32: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Interpolation•Many pixels can not find the corresponding points

because of occluded in the image.

•The corresponding disparity can not be calculated in the disparity after optimized.

• the disparity can be divided into▫matching area▫un-matching area

Page 33: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Interpolation•The disparity in the un-matching area should be

continuity with the disparity in matching area surrounding the un-matching area.

•Disparity interpolation for un-matching area to get the smooth disparity on basis of the optimized disparity.

Page 34: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Interpolation•Step 1 : Mark the matching area and un-matching

▫1 stands for un-matching area.▫0 stands for matching area.

•Step 2 : Find the edge points in the un-matching area▫scanning the 8 neighbourhood of pixels.▫calculate the disparity of edge pixel.▫mark the point as 0.

•Step 3 : Stop scanning until all the points in un-matching area have been matched, or then return to Step2.

Page 35: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Disparity Interpolation

Page 36: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation
Page 37: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Outline• Introduction•Method•Disparity Interpolation•Experiments•Conclusion

Page 38: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Experiments•The results of each steps

Page 39: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation
Page 40: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Experiments•The results of other methods

Page 41: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Experiments•Compare the efficiency

Page 42: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Outline• Introduction•Method•Disparity Interpolation•Proposed Algorithm•Experiments•Conclusion

Page 43: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Conclusion•The dense and smooth disparity can be required.

•The effect of the new algorithm is better than traditional stereo matching algorithm.▫Time complexity can be reduced.▫Efficiency is closer to segment based algorithm.

•The main problem of this algorithm : ▫un-matching area is very big.

Page 44: A  Rapid  S tereo  M atching  A lgorithm  B ased  on  Disparity  I nterpolation

Conclusion•The disparity error of pixels in centre of un-matching

area will be bigger because of interpolating operation is start from the edge of un-matching area.