Dense Optical Flow extraction for GPU and OpenCV 3 rewritten from https://github.com/wanglimin/dense_flow
This tool requires the OpenCV 3.x library to be installed. Make sure that your copy of OpenCV is compiled with CUDA and FFMPEG support. As this is the source of most user errors, here a minimal example on how to compile:
git clone https://github.com/opencv/opencv.git
cd opencv
mkdir build
cd build
cmake -D WITH_CUDA=ON -D WITH_FFMPEG=ON -D CMAKE_INSTALL_PREFIX=/some/path/opencv -D CUDA_ARCH_BIN=61 ..
make -j3
make install
With CUDA_ARCH_BIN you can specify the CUDA compute capability of your card, which typically will speedup compilation. If do not know your compute capability, you can leave that flag out.
To make sure everything worked, check the output of the cmake line and look for CUDA and FFMPEG.
Let's assume OpenCV is installed in /some/path/opencv. You now need to tell dense-flow's Makefile that path. To do this, edit Makefile.config and modify the paths in the second and third line.
Example:
INCLUDE=-I/some/path/opencv/include -Iinclude/
LIB=-L/some/path/opencv/lib
Once this is done, simply run:
mkdir build
make
For convenience, we included a little script build-flow.sh that can batch process videos. You first need to edit line 19 and change LD_LIBRARY_PATH to point to your OpenCV installation. Do NOT forget this step, as the program might otherwise try to load a preinstalled version of OpenCV (which do not have CUDA support). Again assuming OpenCV is installed at /some/path/opencv, you would write:
LD_LIBRARY_PATH=/some/path/opencv/lib build/denseFlow_gpu .........
Now, we need to provide a list of the video files we would like to process. Under Linux, you can do the following:
find /path/to/your/videos -name "*.mp4" > files.txt
This should list all *.mp4 files line-by-line in the textfile files.txt.
Once this is done, you are ready to go. Just execute:
bash build-flow.sh
In case of long video files, generating thousands of jpegs may be inconvenient for copying. We added the possibility to serialize them into one file. If you want to use this feature, please compile with make serialize. Note: Serialization requires boost library!
All jpegs are encoded as a string (using OpenCV's imencode). The resulting vector of string is serialized using boost. unpacker.cc provides an example of how to extract the jpegs.
After 20000 files, the serialized files are written to disk, and a new chunk is started. This can be changed by redefining MAX_FILES_PER_CHUNK in common.hh.
To compile, create 'build' directory, and run 'make serialize'.
Example: ./denseFlow_gpu --vidFile="video.mp4" --xFlowFile="flow_x" --yFlowFile="flow_y" --imgFile="im" --bound=16 --type=2 --device_id=0 --step=10
Reads from file video.mp4, saving the Dense Optical flow with corresponding path prefixes flow_x / flow_y, the images with path prefix im, cuts off all flow values < -16 and > 16, using method 2 (Brox Optical Flow) on GPU 0, sampling only every 10th frame.