Introduction

2021-03-26時点の内容。

Apple Silicon M1でビルドしよう。OpenVINOもNeural Compute Stick 2なら動きます。

Environment

Apple Silicon (M1) MacBook Air
Intel Compute Stick 2 (NCS2)
OpenVINO 2021.3 (No support CPU:arm)
OpenCV 4.5.3-openvino
MacPorts
- sudo port install cmake gflags autoconf automake libtool libusb git-lfs
Miniforge3
- conda install numpy matplotlib cython pandas scipy scikit-learn

build OpenCV 4.5.2-openvino

source tree

git clone --depth 1 -b 4.5.2-openvino https://github.com/opencv/opencv.git opencv-4.5.2-openvino

build

Install dir /opt/intel/openvino/opencv. Same as OpenVINO Distribution.

mkdir opencv-4.5.2-openvino/build
cd opencv-4.5.2-openvino/build

cmake \
    -DCMAKE_INSTALL_PREFIX=/opt/intel/openvino/opencv \
    -DCMAKE_BUILD_TYPE=Release \
    -DWITH_INF_ENGINE=ON \
    -DENABLE_CXX11=ON \
    -DWITH_TBB=ON \
    -DPYTHON3_EXECUTABLE=$HOME/miniforge3/bin/python3.9 \
    -DPYTHON3_LIBRARY=$HOME/miniforge3/lib/libpython3.9.dylib \
    -DPYTHON3_INCLUDE_DIR=$HOME/miniforge3/include/python3.9 ..

make -j 4
sudo make install

build OpenVINO 2021.3

source tree

git clone -b 2021.2 https://github.com/openvinotoolkit/openvino.git openvino-2021.3
cd openvino-2021.3
git submodule update --init --recursive

build

Install dir /opt/intel/openvino. Same as OpenVINO Distribution.

mkdir openvino-2021.3/build
cd openvino-2021.3/build

cmake \
    -DCMAKE_INSTALL_PREFIX=/opt/intel/openvino \
    -DCMAKE_BUILD_TYPE=Release \
    -DENABLE_PYTHON=ON \
    -DPYTHON_EXECUTABLE=$HOME/miniforge3/bin/python3.9 \
    -DPYTHON_LIBRARY=$HOME/miniforge3/lib/libpython3.9.dylib \
    -DPYTHON_INCLUDE_DIR=$HOME/miniforge3/include/python3.9 ..

make -j 4
sudo make install

patch this.

*** setupvars.sh.orig	2020-12-17 12:28:59.000000000 +0900
--- setupvars.sh	2020-12-17 12:29:34.000000000 +0900
***************
*** 18,23 ****
--- 18,24 ----
  BASE_DIR="$( dirname "$SCRIPT_DIR" )"
  
  INSTALLDIR="${BASE_DIR}"
+ INSTALLDIR=/opt/intel/openvino

environment valiable

source /opt/intel/openvino/bin/setupvars.sh
export ngraph_DIR=/opt/intel/openvino/deployment_tools/ngraph/cmake
export OpenCV_DIR=/opt/intel/openvino/opencv 
export PYTHONPATH=$OpenCV_DIR/lib/python3.9/site-packages:$PYTHONPATH

※ お互いに依存しているので OpenCV → OpenVINO →OpenCVでビルド&インストールすること。

Run with Intel Neural Compurte Stick 2

python opencv-4.5.2-openvino/samples/python/video_threaded.py

python /opt/arm64/openvino/deployment_tools/inference_engine/samples/python/classification_sample_async/classification_sample_async.py -m alexnet.xml -d MYRIAD -i image.jpg

OpenCV DNN backend Inference Engine (OpenVINO: Python, C++)

Run Success!! pythonでも動作しています。

Reference

（最小ステップしか書いていないので）ここに書いてあるままでうまく行かなかったり、環境を多少変更したい場合は以下に目を通すとできるんじゃないかな。

TODO

CoreMLも試したい、、、けどねぇ。OpenCV ML backend でCoreMLサポートしないですかね。