neerajgulia/python-opencv-cuda

custom opencv_contrib module which exposes opencv cuda optical flow methods with python bindings

python-opencv-cuda

Solution

1. Create custom opencv_contrib module
2. Write C++ code to wrap the OpenCV CUDA method
3. Using OpenCV python bindings, expose your custom method
4. Build opencv with opencv_contrib
5. Run python code to test

Steps to create the build

Unzip the source:

1. opencv source code: https://github.com/opencv/opencv/archive/3.4.2.zip
2. opencv_contrib source code: https://github.com/opencv/opencv_contrib/archive/3.4.2.zip

Create custom module

1. Copy the folder named "pythoncuda" (inside c++ folder) to: opencv_contrib/modules

Build opencv using following cmake command

1. create build directory inside the opencv folder, cd to the build directory
2. cmake (I used anaconda3 with environment named as: tensorflow_p36 (with python 3.6))

cmake \
-DCMAKE_BUILD_TYPE=RELEASE \
-DWITH_CUDA=ON \
-DCMAKE_INSTALL_PREFIX="/home/$USER/anaconda3/envs/tensorflow_p36" \
-DOPENCV_EXTRA_MODULES_PATH="../../opencv_contrib-3.4.2/modules" \
-DINSTALL_PYTHON_EXAMPLES=OFF \
-DINSTALL_C_EXAMPLES=OFF \
-DBUILD_SHARED_LIBS=OFF \
-DBUILD_DOCS=OFF \
-DBUILD_TESTS=OFF \
-DBUILD_EXAMPLES=OFF \
-DBUILD_PERF_TESTS=OFF \
-DBUILD_opencv_dnn=OFF \
-DTINYDNN_USE_NNPACK=OFF \
-DTINYDNN_USE_TBB=ON \
-DTINYDNN_USE_OMP=ON \
-DENABLE_FAST_MATH=ON \
-DWITH_OPENMP=ON \
-DWITH_TBB=ON \
-DWITH_JPEG=OFF \
-DWITH_IPP=OFF \
-DMKL_WITH_TBB=ON \
-DMKL_WITH_OPENMP=ON \
-DBUILD_opencv_python2=OFF \
-DPYTHON_EXECUTABLE="/home/$USER/anaconda3/envs/tensorflow_p36/bin/python" \
-DPYTHON_LIBRARY="/home/$USER/anaconda3/envs/tensorflow_p36/lib/python3.6" \
-DPYTHON3_LIBRARY="/home/$USER/anaconda3/envs/tensorflow_p36/lib/python3.6" \
-DPYTHON3_EXECUTABLE="/home/$USER/anaconda3/envs/tensorflow_p36/bin/python" \
-DPYTHON3_INCLUDE_DIR="/home/$USER/anaconda3/envs/tensorflow_p36/include/python3.6m" \
-DPYTHON3_INCLUDE_DIR2="/home/$USER/anaconda3/envs/tensorflow_p36/include" \
-DPYTHON3_NUMPY_INCLUDE_DIRS="/home/$USER/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/numpy/core/include" \
-DPYTHON3_INCLUDE_PATH="/home/$USER/anaconda3/envs/tensorflow_p36/include/python3.6m" \
-DPYTHON3_LIBRARIES="/home/$USER/anaconda3/envs/tensorflow_p36/lib/libpython3.6m.so" \
..

2. make
3. sudo make install
4. sudo ldconfig

Test the code

1. Activate conda environment
2. Go to folder: python/ and execute the cpu-opt_flow.py and gpu-opt_flow.py python files

python cpu-opt_flow.py
python gpu-opt_flow.py 

Output at my end:

total time in optical flow CPU processing: 74.15 sec, for: 794 frames. FPS: 10.71

total time in optical flow GPU processing: 21.98 sec, for: 794 frames. FPS: 36.12

Harware configuration:

• CPU - i7 7th Gen
• GPU - NVIDIA TITAN Xp
• RAM - 32 GB

python-opencv-cuda's Issues

Compiled version?

Hello,
I am not proficient in C++ so was wondering if you have the final compiled version that I can use to install in my Python environment somehow?
I tried the steps but they didn't work for me.
Thank you

pyopencv_generated_include.h not found

Thanks for adding this tutorial.

I was able to cmake and make, following your instructions, however the process failed during
sudo make install with the error message:

/home/user/src/opencv-3.4.2/modules/python/src2/cv2.cpp:30:40: fatal error: pyopencv_generated_include.h: No such file or directory
compilation terminated.

This is from one fo your files. Any idea why that file was not generated?

cv2.cv2 has no attribute 'pythoncuda' ?

(DRenv) $USER@pop-os:~/Workspace/FYP/DR/python-opencv-cuda/python$ python cpu-opt_flow.py
Traceback (most recent call last):
File "cpu-opt_flow.py", line 4, in
cm.findOpticalFlow('video/vtest.avi', 'video/cpu_output.avi', False, False)
File "/home/muhdlaziem/Workspace/FYP/DR/python-opencv-cuda/python/common.py", line 63, in findOpticalFlow
flow = cv2.pythoncuda.cpuOpticalFlowFarneback(g_prev, g_next, None, 0.5, 3, 15, 3, 5, 1.2, 0)
AttributeError: module 'cv2.cv2' has no attribute 'pythoncuda'

(DRenv) $USER@pop-os:~/Workspace/FYP/DR/python-opencv-cuda/python$ python gpu-opt_flow.py
Traceback (most recent call last):
File "gpu-opt_flow.py", line 4, in
cm.findOpticalFlow('video/vtest.avi', 'video/cpu_output.avi', True, False)
File "/home/muhdlaziem/Workspace/FYP/DR/python-opencv-cuda/python/common.py", line 60, in findOpticalFlow
flow = cv2.pythoncuda.gpuOpticalFlowFarneback(g_prev, g_next, None, 0.5, 3, 15, 3, 5, 1.2, 0)
AttributeError: module 'cv2.cv2' has no attribute 'pythoncuda'