Efficient High-Resolution Stereo Matching using Local Plane Sweeps Sudipta N. Sinha, Daniel Scharstein, Richard Szeliski @ CVPR 2014 Yongho Shin
𝑂(𝑊𝐻𝑁𝐷) 𝑂( 2 10 𝑊𝐻𝑁𝐷) Problems High-resolution images require long time for computing a disparity map Complexity for general local methods for 2x size images 𝑂(𝑊𝐻𝑁𝐷) 𝑂( 2 10 𝑊𝐻𝑁𝐷) This is a problem of this paper. If we use the high resolution image, it requires a lot of time for calculating disparity map. For example, we have image scaled up 4 times, and we needs almost 1000 times time. x4
Related works 𝑂(𝑊𝐻𝐷) Semi-global matching Optimize following energy function 𝐸=𝐸 𝑑𝑎𝑡𝑎 +𝐸( 𝐷 𝑝 − 𝐷 𝑞 =1)+𝐸( 𝐷 𝑝 − 𝐷 𝑞 >1) NP-hard problem!! Approximate methods operate in adequate computing time, but still slow Dynamic programming gives faster way, but erroneous result Instead do dynamic programming along many directions It cannot model second-order smoothness 𝑂(𝑊𝐻𝐷) For efficiently capturing disparity map, a lot of athors give a good manners. One of the good method is a SGM. SGM is the optimization technique. It minimize following energy function. But this energy function is NP –hard problem.
Related works Efficient large-scale stereo matching Stereo matching based on search space reduction Computation GCPs Delaunay triangulation on GCPs Matching on triangles with restricted range
Segment-Based Stereo Matching Using Belief Propagation Very related work
Matching with a segmentation Initial matching Any matching method can be used Initial matching Extraction of reliable pixels model parameter from each segment Assignment of optimal parameter for each segment by BP Noisy result
Matching with a segmentation Extraction of reliable pixels Simple cross checking method is used Occlusion region can be detected Initial matching Extraction of reliable pixels model parameter from each segment Assignment of optimal parameter for each segment by BP Left image Right image Left result Right result
Matching with a segmentation Extraction of model parameter from each segment At each segment, a model parameter is extracted using reliable pixels and robust statistical technique Add the parameter to a parameter set Initial matching Extraction of reliable pixels model parameter from each segment Assignment of optimal parameter for each segment by BP Reliable pixels Segments
Matching with a segmentation Extraction of model parameter from each segment At each segment, a model parameter is extracted using reliable pixels and robust statistical technique Add the parameter to a parameter set Initial matching Extraction of reliable pixels model parameter from each segment Assignment of optimal parameter for each segment by BP Parameter Parameter Set
Matching with a segmentation Assignment of optimal parameter for each segment by BP Assign an optimal parameter for each segment as total energy can be minimized Initial matching Extraction of reliable pixels model parameter from each segment Assignment of optimal parameter for each segment by BP Parameter #29 Parameter #29 Parameter Set
Matching with a segmentation b c d a : Initial disparity map b : Interpolated result c : Reliable pixel map d : Result from a segmentation
Matching with a segmentation What they did Make plane parameter by segment and initial disparity map Find optimal plane parameters for each segment of the image Select optimal parameters by BP
Proposed method
Information for understanding What they do Make plane parameter by feature points Find optimal plane parameters for each tiles of the image Allowing objects having curved surface Select optimal parameters by SGM
Hypothesis generation Proposed method
Hypothesis generation Feature matching By Harris corner keypoints and upright DAISY descriptors Matching only points along near epipolar line Due to stereo matching But, they allow small vertical misalignments First round Initial set of matches are selected using the ratio test heuristic Second round For obtaining more matched features Horizontal search range is reduced using local estimates
Hypothesis generation Vertical alignment Correct for small vertical misalignments from errors in rectification By fitting a global linear model using RANSAC with matched features 𝑑 𝑦 =𝑎𝑦+𝑏
Hypothesis generation Disparity plane estimation Cluster matched points and find plane parameters Find k number of planes Using variational approach used for mesh simplification Graph based approach with priority queue
Local plane sweeps Proposed method
Local plane sweep Plane for tiles having parallax Because there are curved objects in the world Hence, gives range of ±T pixels of parallax from plane For each plane, investigate similarity among range 2T Optimize by SGM
Local plane sweep Identifying in-range disparities By disparity map, they give cost U AD NCC JUMP
Proposal generation Proposed method
Proposal generation Initial proposals Online proposals Find the planes with associated points within each tile Online proposals Find frequent plane parameter for each tile Propagate the parameter to neighbors
Global optimization Proposed method
Power SGM!! Global optimization We have Plane parameters for each tile Cost U Energy function Power SGM!!
Experiments
Quantitative results
Qualitative results