This is an implementatioin of a stereo matching method described in

@article{Taniai16,
  author    = {Tatsunori Taniai and
               Yasuyuki Matsushita and
               Yoichi Sato and
               Takeshi Naemura},
  title     = {{Continuous Stereo Matching using Local Expansion Moves}},
  journal   = {CoRR (arXiv)},
  volume    = {abs/1603.08328},
  year      = {2016},
  url       = {http://arxiv.org/abs/1603.08328},
}

The code is for research purpose only. If you use our code, please cite the above paper. (Note that we have slightly changed the title in the latest preprint v2 in 2017 from the above v1 in 2016). We also encourage to cite the following conference paper too.

@inproceedings{Taniai14,
  author    = {Tatsunori Taniai and
               Yasuyuki Matsushita and
               Takeshi Naemura},
  title     = {{Continuous Stereo Matching using Locally Shared Labels}},
  booktitle = {{IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}},
  year      = {2014},
  pages     = {1613--1620},
}

See also our project site http://taniai.space/projects/stereo/ and paper https://arxiv.org/abs/1603.08328.

• Visual Studio 2017 Community (installed with the VC++ 2015 vc140 toolset if using the following OpenCV build)
• OpenCV 3 (OpenCV 3.1.0 package will be automatically installed via NuGet upon the initial build)
• Maxflow code v3.01 by Boykov and Kolmogorov (http://vision.csd.uwo.ca/code/)

1. Download and extract maxflow source code to "maxflow" directory. Modify instances.inc to add the following line

template class Graph<float,float,double>;

2. Download and extract an example dataset (see Adirondack below) in "data/MiddV3/Adirondack".
3. Build the solution with release mode. (Doing this will automatically install OpenCV3 package via NuGet. If not, you need to manually install OpenCV3 binaries for the corresponding version of the platform toolset . For the platform toolset vc140, I installed OpenCV by doing "Install-Package opencvcontrib -Version 3.1.0" on the Package Manager console of VS2017.)
4. Run demo.bat file. Results will be saved in "results/cones", "results/teddy", and "results/Adirondack".

• -mode MiddV2: Use settings for Middlebury V2.
• -mode MiddV3: Use settings for Middlebury V3. Assume a MC-CNN matching cost file (im0.acrt) in targetDir.
• -targetDir {string}: Directory that contains target image pairs.
• -outputDir {string}: Directory for saving results. disp0.pfm is the primary result. Intermediate results are also saved in "debug" sub-directory.
• -doDual {0,1}: Estimate left and right disparities and do post-processing using consistency check.
• -iterations {int}: Number of main iterations.
• -pmIterations {int}: Number of initial iterations performed before main iterations without smoothness terms (this accelerates inference).
• -ndisp {int}: Define the disparity range [0, ndisp-1]. It not specified, try to retrieve from files (calib.txt or info.txt).
• -smooth_weight {float}: Smoothness weight (lambda in the paper).
• -filterRedious {int}: The redius of matching windows (ie, filterRedious/2 is the kernel radius of guided image filter).
• -mc_threshold {float}: Parameter tau_cnn in the paper that truncates MC-CNN matching cost values.

• The function of initial iterations (option: pmIterations) is added to accelerate the inference.
• The implementation of guided image filter has been improved from the paper, which reduces the running time of our method by half.

We use matching cost volumes computed by MC-CNN-acrt (https://github.com/jzbontar/mc-cnn). We provide pre-computed matching cost data for 30 test and training image pairs of Middlebury benchmark V3. For demonstration, please use Adirondack below that contains image pairs, calibration data, and ground truth.

• trainingH (11.0 GB): http://www.hci.iis.u-tokyo.ac.jp/datasets/data/LocalExpStereo/trainingH_im0.rar
• testH (7.9 GB): http://www.hci.iis.u-tokyo.ac.jp/datasets/data/LocalExpStereo/testH_im0.rar
• Adirondack (0.6 GB): http://www.hci.iis.u-tokyo.ac.jp/datasets/data/LocalExpStereo/Adirondack.zip

These files only provide cost volume data (im0.acrt) for the left images. The data for the right images can be recovered from the left data (see convertVolumeL2R function in main.cpp). Note that these matching costs are raw outputs from CNNs without cross-based filter and SGM aggregation.