1. Install the Intel® Distribution of OpenVINO™ toolkit core components (latest version: openvino_2019.2.275 that comes with opencv411, without the contrib module and cuda support)

2. Install the dependencies:

• Microsoft Visual Studio* with C++ 2019, 2017, or 2015 with MSBuild

• CMake 3.4 or higher 64-bit

  NOTE: If you want to use Microsoft Visual Studio 2019, you are required to install CMake 3.14.

• Python 3.6.5 64-bit

  IMPORTANT: As part of this installation, make sure you click the option to add the application to your PATH environment variable.

3. Set Environment Variables

   cd C:\Program Files (x86)\IntelSWTools\openvino\bin\

   setupvars.bat

4. Configure the Model Optimizer

   cd C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\model_optimizer\install_prerequisites

   install_prerequisites.bat

5. Run two Verification Scripts to Verify Installation

   cd C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\demo\

   demo_squeezenet_download_convert_run.bat

demo_security_barrier_camera.bat

Caffe model inference test:

    string caffe_root = "C:/Users/FruitLab/OpenCV/OpenCV-Caffe/opencvvino-opencv/caffe-data-images-models/";

    Mat image = imread(caffe_root+"cat.jpg");

    string labels_file = caffe_root+"caffe_ilsvrc12/synset_words.txt";

    string prototxt = caffe_root+"deploy.prototxt";

    string model = caffe_root+"bvlc_reference_caffenet.caffemodel";

    //load labels

    ifstream ifs(labels_file.c_str());

    if(!ifs.is_open())

    {

        CV_Error(Error::StsError, "File" + labels_file + "not found");

        string line;

        while(getline(ifs, line))

        {

            //classes.push_back(line);

        }

    }

    Mat blob;

    blobFromImage(image, blob, 1.0f, Size(224,224), Scalar(104,117,123),false);

    //DNN_BACKEND_INFERENCE_ENGINE

    Net net = readNetFromCaffe(prototxt, model);

    net.setPreferableBackend(DNN_BACKEND_INFERENCE_ENGINE);

    net.setPreferableTarget(DNN_TARGET_CPU);

    cout<<"[INFO] loading model ..."<<endl;

    //prediction

    net.setInput(blob);

    //vector<float> preds = net.forward();

    Mat probs = net.forward();

    double freq = getTickFrequency()/1000;

    vector<double> layersTimes;

    double t = net.getPerfProfile(layersTimes)/freq;

    cout<<"[INFO] classification took "<<t<<" ms"<<endl;

It is noted that without openvino acceleration (just use self-built opencv410 for the task):

About 4-5 times faster due to the use of inference engine!!

References:

http://docs.openvinotoolkit.org/latest/_docs_install_guides_installing_openvino_windows.html#set-the-environment-variables

https://www.learnopencv.com/using-openvino-with-opencv/

https://github.com/spmallick/learnopencv/blob/master/OpenVINO-OpenCV/image-classification.cpp

https://www.pyimagesearch.com/2017/11/06/deep-learning-opencvs-blobfromimage-works/