280 likes | 1.77k Views
Accurate and Efficient Stereo Processing by Semi-Global Matching and Mutual Information. CVPR 05 Heiko Hirschmuller Presenter: Jonghee Park GIST CV-Lab. Introduction. Local stereo matching Local window based method Global stereo matching Dynamic programming Belief propagation Graph-cut
E N D
Accurate and Efficient Stereo Processing by Semi-Global Matching and Mutual Information CVPR 05 HeikoHirschmuller Presenter: JongheePark GIST CV-Lab.
Introduction • Local stereo matching • Local window based method • Global stereo matching • Dynamic programming • Belief propagation • Graph-cut • Semi global matching Streaking in dynamic programming
Semi Global Matching • Global energy minimization problem • D : disparity image, p : point, Dp : disparity of p, Np : neighborhood of p • Penalty setting • Minimizing 2D global minimization is NP-complete for discontinuity preserving energies data term smoothness term (slanted surface) smoothness term (depth discontinuity)
Semi Global Matching • Dynamic programming • 1D optimization in a 2D image • But, suffer from streaking Streaking in dynamic programming
Semi Global Matching • Aggregation • Lr : aggregation path • r : aggregation direction
Semi Global Matching • Aggregation
Semi global matching • Disparity Computation • Left-right consistency check • Median filter
Comparison with different cost SAD + SGM ASW + SGM
Comparison with different cost CENSUS + SGM CENSUS + SGM + LR-Check + Median
Comparison with OpenCV SGBM OpenCV SGBM 16way SGM 8way