Information & Contact

We release mpegflow for easy extraction of MPEG-flow (motion vectors) from video files along with a visualization tool vis, both under the MIT license. We provide Makefiles for Linux / Windows and distribute binary releases.

Please submit bugs on GitHub directly. For any other question, please contact Vadim Kantorov at [email protected] or [email protected].

If you use this code, please cite our work:

@inproceedings{kantorov2014,
      author = {Kantorov, V. and Laptev, I.},
      title = {Efficient feature extraction, encoding and classification for action recognition},
      booktitle = {Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2014},
      year = {2014} }

Below is a visualization of some flow maps of MPI Sintel's alley_1 sequence:

Source	3	15	50
mpegflow
Ground truth

mpegflow

The tool accepts a video file path as command-line argument and writes MPEG-flow (motion vectors) to standard output. By default, the tool outputs the motion vectors arranged in two matrices - dx and dy.

Command-line options:

vis

The tool accepts a video file path and a dump directory as command-line arguments, output of mpegflow on standard input and saveson disk the visualization of motion vectors overlayed on video frames.

Command-line options:

Examples

• To extract motion vectors:

  ./mpegflow examples/mpi_sintel_final_alley_1.avi > examples/alley_1.txt

• To visualize motion vectors:

  mkdir -p examples/vis_dump && ./mpegflow examples/mpi_sintel_final_alley_1.avi | ./vis examples/mpi_sintel_final_alley_1.avi examples/vis_dump

Feel free to copy-paste and run the examples above. More runnable examples are in examples/extract_motion_vectors.sh and examples/vis_motion_vectors.sh. Feel free to use vis.cpp and examples/vis_hue.m as examples of parsing mpegflow output. examples/vis_hue can also be used to produce hue flow visualizations like above.

Building from source

mpegflow depends only on a recent FFmpeg, vis depends on FFmpeg, OpenCV and libpng. The tools are known to work with FFmpeg 3.1 and OpenCV 3.1.

Once the dependencies are visible to g++, run:

make     # to build mpegflow
make vis # to build vis

You will probably end up with a shared build; for a static build, please feel free to play with Makefile.

To build the tools on Windows:

1. Create directory dependencies
2. Extract FFmpeg dev and shared builds to the dependencies directory (for mpegflow and vis)
3. Extract an OpenCV 3.x build to the dependencies directory (for vis)
4. Open VS2015 x64 Native Tools Command Prompt (VS2015 Community Edition will work) and run:

# fix the paths and versions before running

nmake mpegflow.exe FFMPEG_DIR=dependencies\ffmpeg-3.0.1-win64-dev\ffmpeg-3.0.1-win64-dev
# nmake vis.exe OPENCV_DIR=dependencies\opencv-3.1.0\opencv\build\x64\vc14

6. The Windows build is not fully static. You need to keep avutil-54.dll, avformat-56.dll, avcodec-56.dll, swresample-2.dll (for mpegflow) and opencv_world310.dll (for vis) in the same directory as the binary. Note that the instructions and the Makefile assume x64 machine architecture.