This repo is for using GPU to extract videos' dense optical flow fields.
The code is mainly borrowed from Limin Wang's repo.
I've slightly changed the interface to facilitate my own usage.
| Option | Name | Default | Note |
|---|---|---|---|
| f | vidFile | ex.avi | filename of video |
| x | xFlowFile | x | filename prefix of flow x component |
| y | yFlowFile | y | filename prefix of flow x component |
| i | imgFile | i | filename prefix of image |
| b | bound | 20 | specify the maximum (px) of optical flow |
| t | type | 1 | specify the optical flow algorithm |
| d | device_id | 0 | specify gpu id |
| s | step | 1 | specify the step for frame sampling |
| h | height | 0 | specify the height of saved flows, 0: keep original height |
| w | width | 0 | specify the width of saved flows, 0: keep original width |
mkdir flow
./denseFlow_gpu -f ex.avi -x flow/x -y flow/y -i flow/i -b 20 -t 1 -d 0 -s 1 -h 0 -w 0
ex.avi is the input video
flow/ is the folder containing the extracted RGB frames and optical flows
- The RGB frames and optical flows are extracted at the original video resolution (Height, Width, FPS).
- To extract optical flows at a different resolution, I recommend (1) reformat the video to that resolution, (2) extract frames/flows using this repo.
- Options -h & -w controls the resolution of stored frames/flows.
-
Support on OpenCV 3 +
-
Command when installing OpenCV 3 +
cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local -D WITH_CUDA=ON -D WITH_CUBLAS=ON -D WITH_TBB=ON -D WITH_V4L=ON -D WITH_QT=ON -D WITH_OPENGL=ON -D BUILD_PERF_TESTS=OFF -D BUILD_TESTS=OFF -DCUDA_NVCC_FLAGS="-D_FORCE_INLINES" -D WITH_LIBV4L=ON -D CUDA_GENERATION=Auto -D WITH_OPENMP=ON ..
- Changed Commandlineparse on OPENCV:
./denseFlow_gpu -f=ex.avi -x=flow/x -y=flow/y -i=flow/i -b=20 -t=1 -d=0 -s=1 -h=0 -w=0