This repo contains a slight tweak on MXNet version 1.9.1 that enables the ROIAlign operator to take inputs of data type other than just float32. Specifically, accepting uint8 as input is critical to the performance of the compressed video refinement detector.
This build requires OpenCV version 3.4. Follow instructions online to install it, but replace the cmake command with this:
cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local/opencv3.4 -D INSTALL_C_EXAMPLES=ON -D INSTALL_PYTHON_EXAMPLES=ON -D OPENCV_GENERATE_PKGCONFIG=ON -D OPENCV_EXTRA_MODULES_PATH=~/opencv_build/opencv_contrib/modules -D BUILD_EXAMPLES=ON ..
Clone this repo
git clone [email protected]:RyanHTran/mxnet.git
The build configuration is specified in config.cmake. Modify it if necessary for your installation.
Finally, build the package and install Python bindings
mkdir build && cd build
cmake -D OpenCV_DIR=/usr/local/opencv3.4/share/OpenCV/ -D BUILD_CPP_EXAMPLES=0 -D USE_CPP_PACKAGE=0 ..
cmake --build . --parallel 12 && cd .. && pip install -e ./python