Using OpenCV 2.4.4 with FFmpeg in Windows

Question

I know there are other questions dealing with FFmpeg usage in OpenCV, but most of them appear to be outdated.

By opening up the makefiles in CMake, I can verify that I've got the WITH_FFMPEG flag on. My output folder for the OpenCV build contains a bin folder, within which are Debug and Release folders, each containing a copy of a .dll file entitled opencv_ffmpeg244.dll. I can step into the source code of OpenCV when I create a VideoWriter and verify that the function pointers to the .dll get filled correctly. That much appears to be working.

If I use the FOURCC code of CV_FOURCC_PROMPT, the following codecs work properly:

• Microsoft Video 1
• Intel IYUV codec
• Logitech Video (I420)
• Cinepak Codec by Radius
• Full Frames (Uncompressed)

The following codecs do not work properly (ie. produce a 0kb video file):

• Microsoft RLE

If my understanding is correct, using FFMPEG should allow for encoding video using a whole bunch of new codecs (x264, DIVX, XVID, and so on). However, none of these appear in the prompt. Manually setting them by their FOURCC codes using the macro CV_FOURCC(...) also doesn't work. For instance, using this: CV_FOURCC('X','2','6','4') produces the message:

Could not find encoder for codec id 28: Encoder not found

and makes a video file of size 0kb.

Using this: CV_FOURCC('X','V','I','D') produces no error message, and makes a video file of 6kb that will not play in Windows Media Player or VLC.

I tried manually downloaded the Xvid codec from Xvid.org. Once that was installed, it appeared under the VFW selection in the prompt, and the encoding worked properly. So it's close to a solution, but if I try to set the FOURCC code directly, it still fails as above! I have to pick it from the prompt every time. Isn't FFmpeg supposed to include a whole bunch of codecs? If so, why am I manually downloading the codec instead of using the one built into FFmpeg?

What am I missing here? Is there a way to check that FFMPEG is "enabled"? It seems like the only codecs available in the prompt are VFW codecs, not the FFMPEG ones. The .dll has been built and is sitting in the same folder as the executable, but it appears it's not being used in any way.

Lots of related questions here. Hoping to find somebody knowledgeable about the FFmpeg implementation in OpenCV and with some knowledge of how all of these pieces fit together.

Answer 1

how about running ffmpeg and your application separately and transfer images using piped data?

to get video into opencv program,

ffmpeg -i input.mp4 -vcodec mjpeg -f image2pipe -pix_fmt yuvj420p -r 10 -|program.exe

and for recording etc

program.exe|ffmpeg -r 10 -vcodec mjpeg -f image2pipe -i - -vcodec h264 output.mp4

program.exe should be capable of reading concatenated jpeg images from stdin and writing the same to stdout and the above workflow will work. here's some code to read from stdin and display the video.

#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/highgui/highgui.hpp>

using namespace cv;

#if defined(_MSC_VER) || defined(WIN32)  || defined(_WIN32) || defined(__WIN32__) \
    || defined(WIN64)    || defined(_WIN64) || defined(__WIN64__)
# include <io.h>
# include <fcntl.h>
# define SET_BINARY_MODE(handle) setmode(handle, O_BINARY)
#else
# define SET_BINARY_MODE(handle) ((void)0)
#endif

#define BUFSIZE 10240
int main ( int argc, char **argv )
{

    SET_BINARY_MODE(fileno(stdin));
    std::vector<char> data;
    bool skip=true;
    bool imgready=false;
    bool ff=false;
    int readbytes=-1;
    while (1)
    {   
        char ca[BUFSIZE];
        uchar c;
        if (readbytes!=0)
        {
            readbytes=read(fileno(stdin),ca,BUFSIZE);
            for(int i=0;i<readbytes;i++)
            {
                c=ca[i];
                if(ff && c==(uchar)0xd8)
                {
                    skip=false;
                    data.push_back((uchar)0xff);
                }
                if(ff && c==0xd9)
                {
                    imgready=true;
                    data.push_back((uchar)0xd9);
                    skip=true;
                }
                ff=c==0xff;
                if(!skip)
                {
                    data.push_back(c);
                }
                if(imgready)
                {
                    if(data.size()!=0)
                    {
                        cv::Mat data_mat(data);
                        cv::Mat frame(imdecode(data_mat,1));

                        imshow("frame",frame);
                        waitKey(1);
                    }else
                    {
                        printf("warning");
                    }
                    imgready=false;
                    skip=true;
                    data.clear();
                }
            }
        }
        else
        {
            throw std::string("zero byte read");
        }
    }
}    

to write to output something like this should work.

void saveFramestdout(cv::Mat& frame,int compression)
{
    SET_BINARY_MODE(fileno(stdout));
    cv::Mat towrite;
    if(frame.type()==CV_8UC1)
    {
        cvtColor(frame,towrite,CV_GRAY2BGR);
    }else if(frame.type()==CV_32FC3)
    {
        double minVal, maxVal;
        minMaxLoc(frame, &minVal, &maxVal);
        frame.convertTo(towrite, CV_8U, 255.0/(maxVal - minVal), -minVal * 255.0/(maxVal - minVal));
    }
    else{
        towrite=frame;
    }
    std::vector<uchar> buffer;
    std::vector<int> param(2);
    param[0]=CV_IMWRITE_JPEG_QUALITY;
    param[1]=compression;//default(95) 0-100
    imencode(".jpg",towrite,buffer,param);
    uchar* a = &buffer[0];
    ::write(fileno(stdout),a,buffer.size());
}

the problem with the above is the multiple encode/decode of jpegs, which can be partially solved by linking with libjpeg-turbo. or one could go and figure out how to directly pass raw data from ffmpeg and opencv. for my case, this is quite acceptable as most of the overhead is in encoding or the video processing.