OpenPoseのFaceEstimationをwrapperを使わずに書いてみた

概要

openposeのFaceEstimationは、example/openpose/openpose.cppに
wrapperで書かれているがそのままでは、外部から呼び出しにくく、扱いづらい。
簡単な呼び出しで書けないかググったが見つからなかったので、
examples/tutorial_pose/1_extract_from_image.cppを改造し、
実装してみた。

#コード

1_extract_from_image.cpp

DEFINE_string(face_net_resolution,           "368x368", "face net resolution");

~~

    // Step 3 - Initialize all required classes
    op::ScaleAndSizeExtractor scaleAndSizeExtractor(netInputSize, outputSize, FLAGS_scale_number, FLAGS_scale_gap);
    op::CvMatToOpInput cvMatToOpInput{poseModel};
    op::CvMatToOpOutput cvMatToOpOutput;
    op::PoseExtractorCaffe poseExtractorCaffe{poseModel, FLAGS_model_folder, FLAGS_num_gpu_start};
    // faceNetInputSize
    const auto faceNetInputSize = op::flagsToPoint(FLAGS_face_net_resolution, "368x368 (multiples of 16)");
    const auto facenetOutputSize = faceNetInputSize;
    op::FaceExtractorCaffe faceExtractorCaffe{faceNetInputSize, facenetOutputSize, FLAGS_model_folder, FLAGS_num_gpu_start};
    op::PoseCpuRenderer poseRenderer{poseModel, (float)FLAGS_render_threshold, !FLAGS_disable_blending, (float)FLAGS_alpha_pose};

    op::FaceCpuRenderer faceRenderer{(float)FLAGS_render_threshold};
    op::OpOutputToCvMat opOutputToCvMat;
    op::FrameDisplayer frameDisplayer{"OpenPose Tutorial - Example 1", outputSize};
    // Step 4 - Initialize resources on desired thread (in this case single thread, i.e. we init resources here)
    poseExtractorCaffe.initializationOnThread();
    poseRenderer.initializationOnThread();
    faceExtractorCaffe.initializationOnThread();
    faceRenderer.initializationOnThread();

    // ------------------------- POSE ESTIMATION AND RENDERING -------------------------
    // Step 1 - Read and load image, error if empty (possibly wrong path)
    // Alternative: cv::imread(FLAGS_image_path, CV_LOAD_IMAGE_COLOR);
    cv::Mat inputImage = op::loadImage(FLAGS_image_path, CV_LOAD_IMAGE_COLOR);
    if(inputImage.empty())
        op::error("Could not open or find the image: " + FLAGS_image_path, __LINE__, __FUNCTION__, __FILE__);
    const op::Point<int> imageSize{inputImage.cols, inputImage.rows};
    // Step 2 - Get desired scale sizes
    std::vector<double> scaleInputToNetInputs;
    std::vector<op::Point<int>> netInputSizes;
    double scaleInputToOutput;
    op::Point<int> outputResolution;
    std::tie(scaleInputToNetInputs, netInputSizes, scaleInputToOutput, outputResolution)
        = scaleAndSizeExtractor.extract(imageSize);
    // Step 3 - Format input image to OpenPose input and output formats
    const auto netInputArray = cvMatToOpInput.createArray(inputImage, scaleInputToNetInputs, netInputSizes);
    auto outputArray = cvMatToOpOutput.createArray(inputImage, scaleInputToOutput, outputResolution);
    // Step 4 - Estimate poseKeypoints
    poseExtractorCaffe.forwardPass(netInputArray, imageSize, scaleInputToNetInputs);
    const auto poseKeypoints = poseExtractorCaffe.getPoseKeypoints();
    
    //printf("aaa %d %d %d %d\n", inputImage.cols, inputImage.rows, inputImage.type(), CV_8UC3);
    cv::Mat grayImg;
    cvtColor( inputImage, grayImg, CV_BGR2GRAY );
    cv::CascadeClassifier cascade;
    std::string cascadeName = "haarcascade_frontalface_alt.xml";
    cascade.load( cascadeName );
    std::vector<cv::Rect> faces;
    cascade.detectMultiScale( grayImg, faces, 1.1, 2, 0 | CV_HAAR_SCALE_IMAGE, cv::Size(60, 60));
    for( std::vector<cv::Rect>::const_iterator r = faces.begin(); r != faces.end(); r++) {
        //cv::rectangle( inputImage, cvPoint(r->x, r->y), cvPoint(r->x + r->width-1, r->y + r->height-1), CV_RGB(255,0,0), 2, 8, 0);
    }
    //cv::imshow("image", inputImage);![result.png](https://qiita-image-store.s3.amazonaws.com/0/257741/7dda41fc-9a21-4562-085a-0fda3497bce3.png)

    //cv::waitKey(30);
    
    std::vector<op::Rectangle<float> > faceRectangles;
    op::Rectangle<float> fr;
    fr.x = faces[0].x;
    fr.y = faces[0].y;
    fr.width = faces[0].width;
    fr.height = faces[0].height;
    faceRectangles.push_back(fr);
    faceExtractorCaffe.forwardPass(faceRectangles, inputImage);
    const auto face_poseKeypoints = faceExtractorCaffe.getFaceKeypoints();
    
    // Step 5 - Render poseKeypoints
    poseRenderer.renderPose(outputArray, poseKeypoints, scaleInputToOutput);
    faceRenderer.renderFace(outputArray, face_poseKeypoints, scaleInputToOutput);

FaceExtractorCaffeとFaceCpuRendererの定義、初期化を追加し、
顔位置は、opencvで検出して、
検出枠、画像データを入力として、faceExtractorCaffe.forwardPassを実行。
faceRendererでキーポイントを描画している。

#結果

顔部位が検出できました！