mexopencv

Collection and a development kit of MATLAB MEX functions for OpenCV library.

The package provides MATLAB MEX functions that interface a hundred of OpenCV APIs. Also the package contains C++ class that converts between MATLAB's native data type and OpenCV data types. The package is suitable for fast prototyping of OpenCV application in MATLAB, use of OpenCV as an external toolbox in MATLAB, and development of a custom MEX function.

The latest version of mexopencv (master branch) is compatible with OpenCV 3.x. For older OpenCV versions, please checkout the corresponding 2.x branches in (v2.4, v2.3, and v2.1).

Structure
Build
- Unix
- Windows
Usage
Developing a new MEX function
- Testing
- Documenting
License

Structure

The project tree is organized as follows:

+cv/             OpenCV or custom API directory
+mexopencv/      mexopencv utility API directory
doc/             directory for documentation
include/         header files
lib/             directory for compiled C++ library files
samples/         directory for sample application codes
src/             directory for C++ source files
src/+cv/         directory for MEX source files
src/+cv/private/ directory for private MEX source files
test/            directory for test scripts and resources
opencv_contrib/  directory for sources/samples/tests of additional modules
utils/           directory for utilities
Doxyfile         config file for doxygen
Makefile         make script
README.markdown  this file

Build

Prerequisite:

Linux: MATLAB or Octave (>= 4.0.0), OpenCV (>= 3.0.0), g++, make, pkg-config
OS X: MATLAB or Octave (>= 4.0.0), OpenCV (>= 3.0.0), Xcode Command Line Tools, pkg-config
Windows: MATLAB or Octave (>= 4.0.0), OpenCV (>= 3.0.0), supported compiler

Currently, mexopencv targets the final 3.1.0 stable version of OpenCV. You must build it against this exact version, rather than using the bleeding-edge dev-version of opencv or opencv_contrib. UNIX users should consider using a package manager to install OpenCV.

Linux

First make sure you have OpenCV installed in the system. If not, install the package available in your package manager (e.g., libopencv-dev in Debian/Ubuntu, opencv-devel in Fedora), or install the source package from http://opencv.org/ . Make sure pkg-config command can identify OpenCV path. Currently, Ubuntu 16.04 LTS or earlier users have options to:

install OpenCV 3.1 from source and set up PKG_CONFIG_PATH environmental variable to use pkg-config, or
install libopencv-dev apt package and use the v2.4 release of mexopencv.

If you have all the prerequisite, go to the mexopencv directory and type:

$ make

This will build and place all MEX functions inside +cv/. Specify your MATLAB directory if you install MATLAB other than /usr/local/matlab,

$ make MATLABDIR=/opt/local/MATLAB/R2016a

Optionally you can test the library functionality.

$ make test

Developer documentation can be generated with Doxygen if installed.

$ make doc

This will create HTML and LaTeX files under doc/.

Troubleshooting: Invalid MEX file or Segmentation fault

If MATLAB says 'Library not loaded' or any other error in the test, it's likely a compatibility issue between a system library and MATLAB's internal library. You might be able to fix this issue by preloading the library file. On Linux, set the correct library path in LD_PRELOAD environmental variable. For example, if you see GLIBCXX_3.4.15 error in MEX, use the following to start MATLAB.

$ LD_PRELOAD=/usr/lib/libstdc++.so.6 matlab

Note that you need to find the correct path to the shared object. For example, /usr/lib64/ instead of /usr/lib/. You can use locate command to find the location of the shared libraries. On Mac OS X, this environment variable is named DYLD_INSERT_LIBRARIES.

To find what library is conflicting, use ldd command both in the Unix shell and within MATLAB to one of the compiled MEX-files. For example,

$ ldd +cv/imread.mexa64    # within UNIX shell

>> !ldd +cv/imread.mexa64  % within MATLAB

If the output of the ldd command gives you different line, that library is likely to be causing the conflict. Try to preload such a library before launching MATLAB. On Mac, you can use otool -L command instead.

OS X

Currently, the recommended approach to install OpenCV in OS X is Homebrew. Install Homebrew first, and do the following to install OpenCV 3.

$ brew install pkg-config homebrew/science/opencv3
$ brew link opencv3

If you have all the prerequisite, go to the mexopencv directory and type:

$ make MATLABDIR=/Applications/MATLAB_R2016a.app PKG_CONFIG_MATLAB=opencv3 LDFLAGS=-L/usr/local/share/OpenCV/3rdparty/lib -j2

Replace the path to Matlab with your version. This will build and place all MEX functions inside +cv/.

Windows

Refer to this wiki page for detailed instructions on how to compile OpenCV 3.0 with opencv_contrib modules on Windows.

1) Configure a C/C++ compiler for MEX-files in MATLAB

To build mexopencv MEX-files, you need a standard-compliant C++ compiler supported by MATLAB. For an up-to-date list of supported compilers for different versions of MATLAB, see this page.

At the time of writing, Visual Studio 2010 is the recommended version to build mexopencv on Windows platforms. If you are building for a 64-bit target, you have two options:

use Visual Studio Professional edition (make sure "X64 Compilers and Tools" component is chosen during installation)
use Visual C++ Express edition along with latest Windows SDK, both available to download for free.

To select a compiler configuration in MATLAB, type the following command, and follow the instructions (this should be done only once):

>> mex -setup

2) Install OpenCV library

Download the latest pre-built OpenCV binaries from http://opencv.org/downloads.html
Extract/unpack the archive into a destination of your choosing. For example C:\OpenCV
Add the bin folder containing the DLL files to the system PATH environment variable. You should choose the correct binaries depending on your platform and compiler. Example C:\OpenCV\build\x86\vc10\bin. Be careful that the architecture (x86 or x64) should match your MATLAB architecture but not your OS. Also VC version (vc10 or vc11) should match the MEX setup (and probably MATLAB's internal runtime). For example, if you're running MATLAB 32-bit in Windows 7 64-bit with Visual Studio 2010 Express, you should use x86 and vc10. You might need to reboot for changes to take effect.

Alternatively, you can build OpenCV from the sources. Follow this tutorial in the OpenCV documentation for detailed instructions. Just make sure to organize the output in the same directory layout described before, with a structure similar to:

OpenCV
|
+- build
    |-- $ARCH (x86, x64)
    |    |-- $COMPILER (vc10, vc11, vc12, ..)
    |          |-- bin
    |          |    |-- opencv_core300.dll
    |          |    |-- opencv_core300d.dll
    |          |    +-- ...
    |          +-- lib
    |               |-- opencv_core300.lib
    |               |-- opencv_core300d.lib
    |               +-- ...
    +-- include
         |-- opencv
         |    |-- cv.h
         |    +-- ...
         +-- opencv2
              |-- opencv.hpp
              +-- ...

3) Build mexopencv

Once you satisfy the above requirements, you can proceed to build all MEX functions. Browse to mexopencv root folder, and type the following in the MATLAB command window (you need to specify the path where OpenCV library is installed):

>> mexopencv.make('opencv_path', 'C:\OpenCV\build')

Note that if you build OpenCV from source, this path specification might not work out of the box. Follow the directory layout described above to arrange the OpenCV .DLL and .LIB files to correctly compile and link your MEX-files with the library.

To remove existing mexopencv binaries, use the following command.

>> mexopencv.make('clean', true)

Troubleshooting: Invalid MEX-file or Segmentation fault

Check the following common issues first:

Make sure the system PATH is set up correctly. This is different from addpath() in MATLAB. You must have the correct DLL files visible in the system path, depending on the MATLAB architecture and the compiler.
A supported MEX compiler is setup correctly, In Windows 64-bit environment, Windows SDK compiler is needed for Visual C++ Express editions. Note that if you change the compiler configuration, you should clean any previously compiled MEX-files mexopencv.make('clean',true) and build again from scratch.

If you still see the "Invalid MEX-file" error messages, and you are using OpenCV DLL's manually built from sources, check that a consistent value of _SECURE_SCL flag was used during compilation. The current version of mexopencv.make script explicitly adds _SECURE_SCL=1 flag in the build command for Visual Studio compilers older than 2010, so that the built MEX-files are compatible with the OpenCV binary distribution. If you manually built OpenCV with different _SECURE_SCL flag, edit mexopencv.make file and change the flag to use a consistent value.

When unspecified, the default value of the _SECURE_SCL flag depend on the version of the Visual Studio compiler, and whether building is in "Debug" or "Release" mode:

VS2010 and newer: In debug mode, the default value for _SECURE_SCL is 1. In release mode, the default value for _SECURE_SCL is 0.
VS2008 and older: The default value for _SECURE_SCL is 1.

Alternatively, you can change the default value for the _SECURE_SCL flag in mex command. To change the default configuration, which is created with the mex -setup command in MATLAB, is located in the following path in recent versions of Windows.

C:\Users\(Username)\AppData\Roaming\MathWorks\MATLAB\(version)\mexopts.bat

Open this file and edit /D_SECURE_SCL option. Note that this is usually only necessary for VS2008 and older.

If you see "Invalid MEX-file" error even when having matched the _SECURE_SCL flag, it probably indicates some other compatibility issues. Please file a bug report at https://github.com/kyamagu/mexopencv . Specify your OS and compiler, MATLAB and OpenCV versions, along with any error messages and/or crash reports.

Troubleshooting: Visual Studio 2008 compatibility issue

Some users reported incompatibility with Visual Studio 2008. The current recommended version to build mexopencv is VS2010. For this reason, mexopencv on Windows platform does not work with MATLAB R2009b or earlier.

Nevertheless, if you want to try using Visual Studio 2008, obtain stdint.h and use mexopencv.make to compile the package. Visual Studio 2008 or earlier does not comply with C99 standard and lacks stdint.h header file. Luckily, the header file is available on the Web. For example, http://msinttypes.googlecode.com/svn/trunk/stdint.h

Place this file under include directory in the mexopencv package.

Usage

Once MEX functions are compiled, you can add path to the project directory and call MEX functions within MATLAB using package name cv.

addpath('/path/to/mexopencv');
result = cv.filter2D(img, kern);  % with package name 'cv'

import cv.*;
result = filter2D(img, kern);     % no need to specify 'cv' after imported

Note that some functions such as cv.imread overload MATLAB's built-in function when imported. Use the scoped name when you need to avoid name collision. It is also possible to import individual functions. Check help import in MATLAB.

Check a list of functions available by help command in MATLAB.

>> help cv; % shows list of functions in package 'cv'

Contents of cv:

GaussianBlur       - Smoothes an image using a Gaussian filter
Laplacian          - Calculates the Laplacian of an image
VideoCapture       - VideoCapture wrapper class
...

>> help cv.VideoCapture; % shows documentation of VideoCapture

VIDEOCAPTURE  VideoCapture wrapper class

 Class for video capturing from video files or cameras. The class
 provides MATLAB API for capturing video from cameras or for reading
 video files. Here is how the class can be used:
...

Look at the samples/ directory for examples.

Documentation

mexopencv includes a simple documentation utility that generates HTML help files for MATLAB. The following command creates a user documentation under doc/matlab/ directory.

addpath('utils');
MDoc;

On-line documentation is available at http://kyamagu.github.io/mexopencv/ .

Unit Testing

You can test the functionality of compiled files by UnitTest class located inside test directory.

addpath('test');
UnitTest;

Look at the test/unit_tests/ directory for all unit-tests.

Gotchas

OpenCV uses 0-based indexing while MATLAB uses 1-based indexing. That is, the top left pixel is (0,0) in OpenCV whereas MATLAB treats it as (1,1). mexopencv does NOT convert image coordinates. Be careful when accessing a function that deals with image coordinates.
OpenCV often uses channels as dimensions of coordinate representation, as seen in cv.perspectiveTransform. In MATLAB, you can make these channeled array by creating 1xNxd or Nx1xd array for an N-element array of d-dimensional vectors. Hint: use permute or shiftdim functions to convert from/to Nxd numeric array in MATLAB.

Developing a new MEX function

All you need to do is to add your C++ source file in src/+cv/. If you want to add a MEX function called myfunc, create src/+cv/myfunc.cpp. The minimum contents of the myfunc.cpp would look like this:

#include "mexopencv.hpp"
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
    // Check arguments
    nargchk (nlhs<=1 && nrhs==1);

    // Convert MxArray to cv::Mat
    cv::Mat mat = MxArray(prhs[0]).toMat();

    // Do whatever you want

    // Convert cv::Mat back to mxArray*
    plhs[0] = MxArray(mat);
}

This example simply copies an input to cv::Mat object and then copies again to the output. Notice how the MxArray class provided by mexopencv converts mxArray to cv::Mat object. Of course you would want to do something more with the object. Once you create a file, type mexopencv.make() to build your new function. The compiled MEX function will be located inside +cv/ and accessible through cv.myfunc within MATLAB.

The mexopencv.hpp header includes a class MxArray to manipulate mxArray objects. Mostly this class is used to convert between OpenCV data types and mxArray.

int i            = MxArray(prhs[0]).toInt();
double d         = MxArray(prhs[0]).toDouble();
bool b           = MxArray(prhs[0]).toBool();
std::string s    = MxArray(prhs[0]).toString();
cv::Mat mat      = MxArray(prhs[0]).toMat();   // For pixels
cv::Mat ndmat    = MxArray(prhs[0]).toMatND(); // For N-D array
cv::Point pt     = MxArray(prhs[0]).toPoint();
cv::Size siz     = MxArray(prhs[0]).toSize();
cv::Rect rct     = MxArray(prhs[0]).toRect();
cv::Scalar sc    = MxArray(prhs[0]).toScalar();
cv::SparseMat sp = MxArray(prhs[0]).toSparseMat(); // Only double to float

plhs[0] = MxArray(i);
plhs[0] = MxArray(d);
plhs[0] = MxArray(b);
plhs[0] = MxArray(s);
plhs[0] = MxArray(mat);
plhs[0] = MxArray(ndmat);
plhs[0] = MxArray(pt);
plhs[0] = MxArray(siz);
plhs[0] = MxArray(rct);
plhs[0] = MxArray(sc);
plhs[0] = MxArray(sp); // Only 2D float to double

Check MxArray.hpp for the complete list of the conversion API.

If you rather want to develop a MATLAB class that internally calls a MEX function, make use of the +cv/private/ directory. Any function placed under private directory is only accessible from +cv/ directory. So, for example, when you want to design a MATLAB class that wraps the various behavior of the MEX function, define your class at +cv/MyClass.m and develop a MEX function dedicated for that class in src/+cv/private/MyClass_.cpp. Inside of +cv/MyClass.m, you can call MyClass_() without the cv namespace. In mexopencv, this is usually used to exposed C++ classes as MATLAB classes.

Testing

You can optionally add a testing script for your new function. The testing convention in mexopencv is that testing scripts are all written as a static function in a MATLAB class. For example, test/unit_tests/TestFilter2D.m is a class that describes test cases for cv::filter2d function. Inside of the class, a couple of test cases are written as static functions whose name start with 'test'.

If there is such a class inside test/unit_tests/, typing make test would invoke all test cases and show your result. Use test/ directory to place any resource files necessary for testing. An example of testing class is shown below:

classdef TestMyFunc
    methods (Static)
        function test_1
            src = imread(fullfile(mexopencv.root(),'test','img001.jpg'));
            ref = [1,2,3];                  % reference output
            dst = cv.myfunc(src);           % execute your function
            assert(isequal(dst, ref));      % check the output
        end

        function test_error_1
            try
                cv.myfunc('foo');           % myfunc should throw an error
                error('UnitTest:Fail','myfunc incorrectly returned');
            catch e
                assert(strcmp(e.identifier,'mexopencv:error'));
            end
        end
    end
end

In Windows, add the test directory to the MATLAB path and invoke UnitTest to run all the test routines.

Documenting

You can create a MATLAB help documentation for a MEX function by having the same file with '.m' extension. For example, a help file for filter2D.mex* would be filter2D.m. The help file should only contain MATLAB comments. An example is shown below:

%MYFUNC  brief description about myfunc
%
%    out = cv.myfunc(in)
%
% ## Input
% * __in__ input image.
%
% ## Output
% * __out__ output image.
%
% Detailed description of function continues...
%

License

The code may be redistributed under the BSD 3-Clause license.

hemprasad/mexopencv