I Hate (Love) OpenCV

Things that just don't work

2006-04-24T01:30:00.000-07:00

An ongoing list:

cvMinMaxLoc on non-32F images
cvPyrUp / cvPyrDown on 32F images (works on 8U)

Why? Who knows.

Getting Started with OpenCV in Visual Studio 2005

2006-04-20T02:57:00.000-07:00

Step 1: Create a new project.

Step 2: Select Win32 Console Application

Step 3: We want an Empty project

Step 4: Done creating the project

This is what your project screen should look like at this point.

Step 5: Create a new item

Step 6: Create a C++ file, the name is your choice.

Step 7: Modify the project properties

Step 7: Select All Configurations

Step 9: Go to Configuration Properties / C/C++ / General / Additional Include Directories

Step 10: Enter the path to your include files: For default OpenCV install location, they are

"C:\Program Files\OpenCV\cvaux\include\";"C:\Program Files\OpenCV\cxcore\include\";"C:\Program Files\OpenCV\cv\include\";"C:\Program Files\OpenCV\otherlibs\highgui\";"C:\Program Files\OpenCV\otherlibs\cvcam\include\"

Step 10: Enter the path to your include files: For default OpenCV install location, they are

"C:\Program Files\OpenCV\cvaux\include\";"C:\Program Files\OpenCV\cxcore\include\";"C:\Program Files\OpenCV\cv\include\";"C:\Program Files\OpenCV\otherlibs\highgui\";"C:\Program Files\OpenCV\otherlibs\cvcam\include\"

Step 11: Go to Configuration Properties / Linker / Input / Additional Dependencies

Enter the path to all the relevant cv libs. For a default install, this is:

"C:\Program Files\OpenCV\lib\cv.lib" "C:\Program Files\OpenCV\lib\cxcore.lib" "C:\Program Files\OpenCV\lib\cvaux.lib" "C:\Program Files\OpenCV\lib\highgui.lib"

(notice the spaces instead of ; marks)

Write some code

Click on "Start Debug"

Hello World in OpenCV

cvReleaseImage() on image with ROI set

2005-04-16T19:35:00.000-07:00

This one took me a long time to track down, so make sure you clear an image's ROI before you call cvReleaseImage() on it. The error I kept getting (which of course is very informative) was:


OpenCV ERROR: Unknown error code -49 (Deallocation error)
      in function cvFree, cxalloc.cpp(136)
Terminating the application...

cvIntegral() on 32-bit floating point images

2005-04-16T16:02:00.001-07:00

OpenCV provides a very useful function performing fast Viola and Jones object detection, but it can be a little silly in its requirements for the target image format. For the most part, it insists on you using double precision floating point integral images, and if you need to store a whole bunch of these in memory, you can quickly run out of RAM even on well equipped modern machines. Not finding an easy way to force it to use floats instead of doubles, I wrote my own integral image function, which frankly, looks simpler than OpenCV's, which you are free to use:

http://www.etalonsoft.com/code/integral.cpp

Montage in OpenCV

2005-04-05T00:22:00.000-07:00

MATLAB has a really convenient function to create a montage of images -- i.e. a large image from a whole bunch of little ones. Requiring this functionality in my OpenCV projects, I decided to write my own utility function that accomplishes the same goal. Here's the code for it, and below is an example montage image produced:

Example montage image with 10 images per row

OpenCV Assertions Are Lame

2005-03-19T23:22:00.000-08:00

Whenever OpenCV library functions encounter a problem such as a NULL pointer passed in, or unexpected dimensions, or unexpected image format, they croak right away and for your help, apparently, output a line number deeeep inside the function framework where the problem occurred, which is often of little use to you. Here's an example:

OpenCV ERROR: Null pointer (NULL array pointer is passed)
in function cvGetMat, cxarray.cpp(2766)
Terminating the application...

If you're compiling on the command line, you'll have to bust out gdb and hope you're a gdb pro to find what high-level function call in your code causes the assertion to fail.

Switch to C++

2005-03-17T16:17:00.000-08:00

I've decided to switch to C++, primarily because I like to use default parameters. For example, in the OpenCV 0.9.6 Documentation the Canny edge detector is defined as:


void cvCanny( const CvArr* image, CvArr* edges, double threshold1,
              double threshold2, int aperture_size=3 );

That means that you can call cvCanny like this:


cvCanny(sourceimage, resultimage, low_threshold, high_threshold);

Or like this:


cvCanny(sourceimage, resultimage, low_threshold, high_threshold, aperature_size);

And either way will work in C++. C does not support this functionality. So, from now on, anything I post here probably won't compile like this:


gcc test.cpp -o test `pkg-config --libs opencv --cflags opencv`

But it will compile like this:


g++ test.cpp -o test `pkg-config --libs opencv --cflags opencv`

Even though the underlying program may be entirely C except for the OpenCV calls.

Documentation

2005-03-17T16:06:00.000-08:00

Earlier, I had complained that the documentation for OpenCV that I could find easily was out of date. I was really looking for some kind of online reference, but it turns out the most up to date documentation is bundled with the distribution. It lives in a subfolder called (surprise) "docs". When you install OpenCV it puts these files by default in the folder /usr/local/share/opencv/doc/.

Opening the file index.htm gives you links to four reference files (OpenCV is really four separate libraries bundled together). The problem is, on my Linux machine, the links don't work. They don't work because the names of the files referenced by the link have capitals, where the actual filenames they reference are all lower case. I fixed this on my local machine and then decided I should put this fixed version online where people could get to it:

OpenCV 0.9.6 Documentation

Hopefully, this will allow for quick and easy reference (and as soon as google indexes it, a convenient site search).

My First OpenCV Program

2005-03-08T18:13:00.000-08:00

Now that I've gotten OpenCV installed, the first thing I needed to try was to create a sample program of some kind that was basically not useful for anything except to demonstrate that I could compile an OpenCV program. Heres the code for test.c, my first attempt:


#include "cv.h"
#include "highgui.h"

int main( int arc, char ** argv)
{
    int W = 640;
    int H = 480;

    // Images are stored in IplImage structs. The function cvCreateImage
    // allocates space for an image of a certain type, in this case an
    // 8 bit, 1 channel (grayscale) image.
    IplImage *image = cvCreateImage(cvSize(W, H), IPL_DEPTH_8U, 1);

    // Create a named window and show our image on it
    cvNamedWindow("My First OpenCV Program", 1);
    cvShowImage("My First OpenCV Program", image);

    // Wait for any keystroke and then release the image from memory and exit
    // the program
    cvWaitKey(0);
    cvReleaseImage(&image);
    return 0;
}

All this code does is create a window and show an image of size 640x480 that was created using cvCreateImage. The image is black, which isn't surprising since we didn't put anything in it. In order to compile it from a shell I did:


gcc test.c -o test `pkg-config --libs opencv --cflags opencv`

How I Installed OpenCV

2005-03-08T00:08:00.000-08:00

Thought I would start things off by recounting my installation experience. I'm running a pretty much default installation of Fedora Core 3 (I did install all the development packages, as well as gtk+2). I downloaded the Linux tar.gz from Sourceforge into my home directory and then from a shell did:


tar -xzf opencv-0.9.6.tar.gz
cd opencv-0.9.6
./configure
make
su -
make install
exit (to get out of root)

This all completed without any real trouble, and things looked good. Problem was, after this step I couldn't compile any opencv programs. One method for compiling an OpenCV program test.c is to type:


gcc test.c -o test `pkg-config --libs opencv --cflags opencv`

When I did this, pkg-config reported that it couldn't find opencv and that perhaps I should add opencv.pc to the PKG_CONFIG_PATH environment variable. I instead decided to add /usr/local/lib (the location of opencv.pc) to my /etc/ld.so.conf file and then run ldconfig (as root) to rebuild the links. You can check to see if everything is working by typing the following at the shell:


pkg-config --libs opencv --cflags opencv

which in my case returned:


-I/usr/local/include/opencv -L/usr/local/lib -lcxcore -lcv -lhighgui -lcvaux

If you get this far, chances are you have everything you need to compile an opencv program except the source code.

Welcome

2005-03-07T23:57:00.000-08:00

I've created this blog because I started experimenting with OpenCV and looking into it's enormous potential only to notice that the documentation is ... lacking. Even Google searches turn up less than impressive results, usually returning one of two sites:

University of Birmingham
A lecture with sample code
More comprehensive documentation

The Official Sourceforge Page
http://sourceforge.net/projects/opencvlibrary/

Admittedly, I've seen less documentation, but never on a project so large and well known. Even the documentation that exists, such as the sample code in the above link, doesn't compile without some modifications to the source. So as I run into problems and hopefully find solutions I'll post results here.