如何从 ML Kit 人脸标志点估计人脸姿势

Posted 2023-03-16

【中文标题】如何从 ML Kit 人脸标志点估计人脸姿势

【英文标题】：How to estimate face pose from ML Kit face landmark points

【发布时间】：2019-08-23 05:27:30

【问题描述】：

我在 ios 上使用 Firebase ML Kit 来检测人脸。虽然它提供了欧拉 Y 角和 Z 角，但它不提供欧拉 X 角（螺距）。所以我想尝试使用 OpenCV solvePnp 计算音高，如下所述： https://www.learnopencv.com/head-pose-estimation-using-opencv-and-dlib/#code

这是我的 Objective C 函数：

+(void) estimatePose:(FIRVisionFace *)face imgSize:(CGSize)imgSize 

    // Contour legend: https://firebase.google.com/docs/ml-kit/images/examples/face_contours.svg
    FIRVisionFaceContour* faceOval = [face contourOfType:FIRFaceContourTypeFace];
    FIRVisionFaceContour* leftEyeContour = [face contourOfType:FIRFaceContourTypeLeftEye];
    FIRVisionFaceContour* rightEyeContour = [face contourOfType:FIRFaceContourTypeRightEye];
    FIRVisionFaceContour* noseBridge = [face contourOfType:FIRFaceContourTypeNoseBridge];
    FIRVisionFaceContour* upperLipTop = [face contourOfType:FIRFaceContourTypeUpperLipTop];

    FIRVisionPoint* chin = faceOval.points[18];
    FIRVisionPoint* leftEyeLeftCorner = leftEyeContour.points[0];
    FIRVisionPoint* rightEyeRightCorner = rightEyeContour.points[8];
    FIRVisionPoint* noseTip = noseBridge.points[1];
    FIRVisionPoint* leftMouthCorner = upperLipTop.points[0];
    FIRVisionPoint* rightMouthCorner = upperLipTop.points[10];

    // 2D/3D model points using https://www.learnopencv.com/head-pose-estimation-using-opencv-and-dlib/#code
    image_points.push_back( cv::Point2d(noseTip.x.doubleValue, noseTip.y.doubleValue) );    // Nose tip
    image_points.push_back( cv::Point2d(chin.x.doubleValue, chin.y.doubleValue) );    // Chin
    image_points.push_back( cv::Point2d(leftEyeLeftCorner.x.doubleValue, leftEyeLeftCorner.y.doubleValue) );    // Left eye left corner
    image_points.push_back( cv::Point2d(rightEyeRightCorner.x.doubleValue, rightEyeRightCorner.y.doubleValue) );     // Right eye right corner
    image_points.push_back( cv::Point2d(leftMouthCorner.x.doubleValue, leftMouthCorner.y.doubleValue) );    // Left Mouth corner
    image_points.push_back( cv::Point2d(rightMouthCorner.x.doubleValue, rightMouthCorner.y.doubleValue) );    // Right mouth corner

    model_points.push_back(cv::Point3d(0.0f, 0.0f, 0.0f));               // Nose tip
    model_points.push_back(cv::Point3d(0.0f, -330.0f, -65.0f));          // Chin
    model_points.push_back(cv::Point3d(-225.0f, 170.0f, -135.0f));       // Left eye left corner
    model_points.push_back(cv::Point3d(225.0f, 170.0f, -135.0f));        // Right eye right corner
    model_points.push_back(cv::Point3d(-150.0f, -150.0f, -125.0f));      // Left Mouth corner
    model_points.push_back(cv::Point3d(150.0f, -150.0f, -125.0f));       // Right mouth corner

    double focal_length = imgSize.width; // Approximate focal length.
    cv::Point2d center = cv::Point2d(imgSize.width / 2, imgSize.height / 2);
    cv::Mat camera_matrix = (cv::Mat_<double>(3,3) << focal_length, 0, center.x, 0 , focal_length, center.y, 0, 0, 1);
    cv::Mat dist_coeffs = cv::Mat::zeros(4,1,cv::DataType<double>::type); // Assuming no lens distortion

    // Output rotation and translation
    cv::Mat rotation_vector; // Rotation in axis-angle form
    cv::Mat translation_vector;

    // Solve for pose
    cv::solvePnP(model_points, image_points, camera_matrix, dist_coeffs, rotation_vector, translation_vector);

    NSLog(@"Rotation Vector %f %f %f", rotation_vector.at<float>(0), rotation_vector.at<float>(1), rotation_vector.at<float>(2));

函数末尾的 print 语句正在为大部分笔直的摆脸产生不切实际的值，例如：

Rotation Vector 430085088834445557973284941130104832.000000 -2.169656 -37005999085886043916656699445870592.000000
Rotation Vector -31942776355636779193640943616.000000 -2.163979 51918482290645906067188728866013184.000000

我做错了什么？

【参考方案1】：

问题在于日志语句。它需要打印双精度而不是浮点数：

NSLog(@"Rotation Vector %f %f %f", rotation_vector.at<double>(0), rotation_vector.at<double>(1), rotation_vector.at<double>(2));

另外，旋转向量不是欧拉角。需要转换：

http://answers.opencv.org/question/16796/computing-attituderoll-pitch-yaw-from-solvepnp/?answer=52913#post-id-52913

【讨论】：

嗨 Jacob，我有兴趣将 EulerX 检测添加到 ML Kit Vision 的颤振库中。您的解决方案效果如何？