• Acta Optica Sinica
  • Vol. 38, Issue 2, 0215006 (2018)
Hairui Fan1、2, Fan Yang1、2、*, Xuran Pan1、2, Jie Wen1、2, and Xiaoyu Wang1、2
Author Affiliations
  • 1 School of Electronic and Information Engineering, Hebei University of Technology, Tianjin 300401, China
  • 2 Tianjin Key Laboratory of Electronic Materials and Devices, Tianjin 300401, China
  • show less
    DOI: 10.3788/AOS201838.0215006 Cite this Article Set citation alerts
    Hairui Fan, Fan Yang, Xuran Pan, Jie Wen, Xiaoyu Wang. Stereo Matching Algorithm for Improved Census Transform and Gradient Fusion[J]. Acta Optica Sinica, 2018, 38(2): 0215006 Copy Citation Text show less
    Support window coordinates and weight distribution map with size of 3×3 and standard deviation of 1.5. (a) coordinate distribution map; (b) weight coordinate distribution map; (c) weight distribution map
    Fig. 1. Support window coordinates and weight distribution map with size of 3×3 and standard deviation of 1.5. (a) coordinate distribution map; (b) weight coordinate distribution map; (c) weight distribution map
    Disparity maps before and after improvement by guidance filter algorithm. (a)(c) before improvement; (b)(d) after improvement
    Fig. 2. Disparity maps before and after improvement by guidance filter algorithm. (a)(c) before improvement; (b)(d) after improvement
    Disparity maps obtained by different matching cost algorithms. (a) CT; (b) MCT; (c) GRD;(d) proposed algorithm
    Fig. 3. Disparity maps obtained by different matching cost algorithms. (a) CT; (b) MCT; (c) GRD;(d) proposed algorithm
    Disparity maps obtained by different aggregation algorithms. (a) BoxF, R=10.81%; (b) BF, R=8.12%;(c) GF, R=7.85%; (d) MST, R=8.31%; (e) proposed algorithm, R=7.57%
    Fig. 4. Disparity maps obtained by different aggregation algorithms. (a) BoxF, R=10.81%; (b) BF, R=8.12%;(c) GF, R=7.85%; (d) MST, R=8.31%; (e) proposed algorithm, R=7.57%
    Comparison of false matching rates of different cost aggregation algorithms in no-occluded regions
    Fig. 5. Comparison of false matching rates of different cost aggregation algorithms in no-occluded regions
    Test results of different stereo matching algorithms on Aloe image pairs. (a) Aloe left image;(b) Aloe right image; (c) GRD, R=10.19%; (d) MCT, R=10.07%; (e) MCT', R=9.74%; (f) proposed algorithm, R=6.74%
    Fig. 6. Test results of different stereo matching algorithms on Aloe image pairs. (a) Aloe left image;(b) Aloe right image; (c) GRD, R=10.19%; (d) MCT, R=10.07%; (e) MCT', R=9.74%; (f) proposed algorithm, R=6.74%
    Test results of different stereo matching algorithms on Baby1 image pairs. (a) Baby1 left image; (b) Baby1 right image; (c) GRD, R=12.82%; (d) MCT, R=4.91%; (e) MCT', R=4.53%; (f) proposed algorithm, R=3.75%
    Fig. 7. Test results of different stereo matching algorithms on Baby1 image pairs. (a) Baby1 left image; (b) Baby1 right image; (c) GRD, R=12.82%; (d) MCT, R=4.91%; (e) MCT', R=4.53%; (f) proposed algorithm, R=3.75%
    Test results of different stereo matching algorithms on Bowling2 image pairs. (a) Bowling2 left image; (b) Bowling2 right image; (c) GRD, R=12.91%; (d) MCT, R=15.77%; (e) MCT', R=14.21%; (f) proposed algorithm, R=9.04%
    Fig. 8. Test results of different stereo matching algorithms on Bowling2 image pairs. (a) Bowling2 left image; (b) Bowling2 right image; (c) GRD, R=12.91%; (d) MCT, R=15.77%; (e) MCT', R=14.21%; (f) proposed algorithm, R=9.04%
    Test results of different stereo matching algorithms on Dolls image pairs. (a) Dolls left image; (b) Dolls right image; (c) GRD, R=7.66%;(d) MCT, R=11.55%; (e) MCT', R=11.87%; (f) proposed algorithm, R=5.77%
    Fig. 9. Test results of different stereo matching algorithms on Dolls image pairs. (a) Dolls left image; (b) Dolls right image; (c) GRD, R=7.66%;(d) MCT, R=11.55%; (e) MCT', R=11.87%; (f) proposed algorithm, R=5.77%
    Experimental results of proposed algorithm on Middlebury2.0 image pairs. (a) Testing left image; (b) standard disparity map; (c) disparity map of proposed algorithm(without disparity refinement); (d) mismatched map (without disparity refinement); (e) disparity maps obtained by proposed algorithm (disparity refinement); (f) mismatched map (disparity refinement)
    Fig. 10. Experimental results of proposed algorithm on Middlebury2.0 image pairs. (a) Testing left image; (b) standard disparity map; (c) disparity map of proposed algorithm(without disparity refinement); (d) mismatched map (without disparity refinement); (e) disparity maps obtained by proposed algorithm (disparity refinement); (f) mismatched map (disparity refinement)
    ParameterωkTgεσThω1TcenλGω2
    Value94.3351×10-51.510.9450.0650250.1
    Table 1. Parameters involved in proposed stereo matching algorithm
    AlgorithmSADCTMCTMCT’GRDProposed algorithm
    Tsukuba4.653.544.584.502.722.53
    Venus3.452.003.633.551.681.60
    Teddy14.218.5613.0511.327.457.57
    Cones7.704.877.496.784.454.04
    Avg7.504.747.196.544.083.93
    Table 2. Percentage of false match in no-occluded region of different matching cost algorithms%
    AlgorithmSADCTMCTMCT’GRDProposed algorithm
    Tsukuba4.973.123.363.593.592.86
    Venus8.345.257.177.154.124.13
    Teddy25.5419.2823.4822.7617.5617.27
    Cones23.4017.0419.7118.8816.1316.06
    Avg15.5611.1713.4313.1010.3510.08
    Table 3. Percentage of false match in all regions of different matching cost algorithms%
    AlgorithmTsukubaVenusTeddyConesAverage
    No-occludedAllNo-occludedAllNo-occludedAllNo-occludedAll
    BPcompressed[21]2.683.631.331.898.3613.93.719.855.67
    GC-occ[5]1.192.011.642.1911.2017.405.3612.46.67
    AdaptAggrDP[22]1.573.501.532.696.7914.35.5313.26.14
    RTCensus[23]5.086.251.582.427.9613.84.109.546.34
    FastAggreg[24]1.162.114.034.759.0415.25.3712.66.78
    SemiGlob[25]3.263.961.001.576.0212.23.069.755.10
    GradAdaptWgt[26]2.262.636.991.398.0013.102.617.675.58
    Proposed algorithm2.523.230.231.055.7012.353.299.754.74
    Table 5. False matching rates of different stereo matching algorithms in no-occlusion region and all regions%
    StereopairsFalse matching rate /%
    MCT[16]MCT'[17]MST[6]AW[13]GF[14]CT-GF[15]CT-MST[15]Proposed algorithm
    AloeArtBaby1Baby2Baby3BooksBowling1Bowling2Cloth1Cloth2Cloth3Cloth4DollsFlowerpotsLampshade1Lampshade2LaundryMidd1Midd2MoebiusMonopolyPlasticReindeerRocks19.4918.614.296.026.9912.2618.2211.151.985.794.053.5811.4314.2724.7327.6024.7748.7547.7014.1628.4944.4013.835.389.3119.213.915.986.7211.3417.2311.161.875.214.143.6412.1515.2125.4327.7725.2140.1440.0113.8627.9739.4312.445.196.6113.747.6514.129.5312.9719.1412.631.185.403.202.537.8617.2711.2914.2019.3918.9620.2111.3316.8530.3011.964.146.5012.885.4513.648.7612.6516.8710.431.255.563.813.216.4315.4412.3617.4416.8936.4735.5613.2615.4332.4411.735.437.0612.033.163.974.7110.216.218.181.884.202.792.257.129.6312.1212.4820.6436.9534.0210.6419.7427.067.843.977.0412.163.314.034.979.966.188.471.964.832.802.387.279.6710.9310.5418.5433.4532.9610.9218.9922.228.154.036.5713.867.9113.679.4712.4719.5812.581.155.363.162.407.6217.0511.3714.0817.8619.2119.3211.6616.6431.0611.923.975.9310.162.532.663.918.016.075.881.243.432.291.645.267.5310.0411.1615.5131.7631.909.5719.5025.806.892.86
    Rocks2Wood1Wood2Average4.136.494.4115.673.247.174.3714.793.3311.174.2011.525.028.996.1512.592.674.932.7110.342.744.932.959.873.9810.984.1211.451.943.722.188.67
    Table 6. False matching rate of different stereo matching algorithms in no-occluded region
    MethodMCT[16]MCT'[17]MST[6]AW[13]GF[14]CT-GF[15]CT-MST[15]Proposed
    Averagerunning time /s3.974.351.5615.874.766.754.126.63
    Table 7. Average running time of 31 Middlebury stereo pairs with different algorithmss
    Hairui Fan, Fan Yang, Xuran Pan, Jie Wen, Xiaoyu Wang. Stereo Matching Algorithm for Improved Census Transform and Gradient Fusion[J]. Acta Optica Sinica, 2018, 38(2): 0215006
    Download Citation