KZ2 - Kolmogorov and Zabih’s graph cuts stereo matching algorithm
by Vladimir Kolmogorov <vnk@cs.cornell.edu>
and Pascal Monasse <monasse@imagine.enpc.fr>

This software is linked to the IPOL article [1], which gives a detailed description of the algorithm. The algorithm was first published in [2]. The source code is an extensively edited version of [3] by Pascal Monasse and is distributed under the terms of the GPLv3 license. The max-flow algorithm is described in [4].

[1] "Kolmogorov and Zabih’s Graph Cuts Stereo Matching Algorithm"
by Vladimir Kolmogorov, Pascal Monasse, Pauline Tan (2014)
    IPOL, http://www.ipol.im/pub/pre/97/
[2] "Computing Visual Correspondence with Occlusions using Graph Cuts"
by Vladimir Kolmogorov and Ramin Zabih (2001)
    International Conference on Computer Vision (IJCV)
[3] "Match 3.4"
software written by Vladimir Kolmogorov <vnk@cs.cornell.edu>
    http://pub.ist.ac.at/~vnk/software.html
[4] "An Experimental Comparison of Min-Cut/Max-Flow Algorithms for Energy Minimization in Vision."
by Yuri Boykov and Vladimir Kolmogorov (2004)
    IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)

Version 1.0 released on 2014/10/10

Future releases and updates:
https://github.com/pmonasse/disparity-with-graph-cuts.git

Build
-----
Prerequisites: CMake version 2.6 or later

- Unix, MacOS, Windows with MinGW:
$ cd /path_to_KZ2/
$ mkdir Build && cd Build && cmake -DCMAKE_BUILD_TYPE:bool=Release ..
$ make

CMake tries to find libPNG and libTIFF on your system, though neither is mandatory. They need to come with header files, which are provided by "...-dev" packages under Linux Debian or Ubuntu. If not found, they are compiled from scratch (source code in src/third_party).

- Windows with Microsoft Visual Studio (MSVC):
1. Launch CMake, input as source code location the KZ2 folder and use a new folder for binaries.
2. Press 'Configure' then 'Generate' buttons. Choose your MSVC version. When done, you can quit CMake.
3. In the binary folder, open the solution file KZ2.sln, this launches MSVC. Select 'Release' mode, then 'Build solution' (F7) will build your program.

- Test:
$ bin/KZ2 ../images/scene_l.png ../images/scene_r.png -15 0 disp.tif -o disp.png
Input images must be epipolar rectified, the disparity range depends on the pair. The output images are in this test:
- disp.tif (float TIFF image, able to contain negative values and with occluded pixels as NaN, Not A Number) and
- disp.png, representing the same image directly viewable, with gray levels for disparity and cyan color for occluded pixels.
The latter is useful as most image viewers do not understand float TIFF.
The file disp.png should be similar to the one in folder ../images but may be slightly different, due to the random order of alpha.

Usage
-----
bin/KZ2 [options] im1.png im2.png dMin dMax [dispMap.tif]
General options:
 -i,--max_iter iter: max number of iterations
 -o,--output disp.png: scaled disparity map
 -r,--random: random alpha order at each iteration
Options for cost:
 -c,--data_cost dist: L1 or L2
 -l,--lambda lambda: value of lambda (smoothness)
 --lambda1 l1: smoothness cost not across edge
 --lambda2 l2: smoothness cost across edge
 -t,--threshold thres: intensity diff for 'edge'
 -k k: cost for occlusion
If no output is given (neither dispMap.tif nor -o option), the program just displays the recommended computed values for K and lambda.

Files
-----
Only files with (*) are reviewed in the IPOL publication.
README.txt
CMakeLists.txt
images/scene_l.png
images/scene_r.png
src/CMakeLists.txt
src/cmdLine.h
src/io_tiff.h
src/io_tiff.c
src/io_png.h
src/io_png.c
src/nan.h
src/image.h (*)
src/image.cpp (*)
src/kz2.cpp (*)
src/match.h (*)
src/match.cpp (*)
src/data.cpp (*)
src/statistics.cpp (*)
src/main.cpp (*)
src/energy/energy.h (*)
src/energy/test_energy.cpp
src/maxflow/graph.h
src/maxflow/graph.cpp
src/maxflow/maxflow.cpp
src/third_party/... (sources of libPNG, libTIFF and their dependencies)

Limitations
-----------
The software is a bit slower (10-20%) than the original code Match of V. Kolmogorov due to memory management of the graph. Match allocates sets of nodes and edges with malloc/realloc, and stores directly pointers to link nodes and edges. It has thus to adjust the pointers when realloc changes array address. This results in ugly code with offsets to pointers but:
- Match has faster max-flow computation since it uses pointers to follow paths while KZ2 uses index in std::vector.
- It was noticed that with the same allocation policy, using C's alloc/realloc is faster than standard allocator of std::vector using C++'s new. The reason is a mystery since elements have no constructor/destructor.
To alleviate the latter defect, a preset amount of memory is pre-allocated for node and edge arrays: 2n nodes and 12n edges, with n the number of pixels (see Match::ExpansionMove in kz2.cpp). These are the maximum possible values, but this pre-allocation is less elegant and less efficient than on-demand allocation.