184 lines
6.9 KiB
C++
184 lines
6.9 KiB
C++
#include "normalizer.hpp"
|
|
|
|
using namespace cv;
|
|
using namespace std;
|
|
|
|
namespace opengaze {
|
|
|
|
Normalizer::Normalizer() {
|
|
// parameters for data normalization
|
|
focal_norm_ = 1600;
|
|
distance_norm_ = 1000; // 600 500 1000
|
|
roiSize_norm_ = cv::Size(224, 224); // 224 448
|
|
cam_norm_ = (Mat_<float>(3,3) << focal_norm_, 0, roiSize_norm_.width/2, 0, focal_norm_, roiSize_norm_.height/2.0f, 0, 0, 1.0f);
|
|
}
|
|
|
|
Normalizer::~Normalizer() {}
|
|
|
|
void Normalizer::setParameters(int focal_length, int distance, int img_w, int img_h){
|
|
// parameters for data normalization
|
|
focal_norm_ = focal_length;
|
|
distance_norm_ = distance; // 600 500 1000
|
|
roiSize_norm_ = cv::Size(img_w, img_h); // 224 448
|
|
cam_norm_ = (Mat_<float>(3,3) << focal_norm_, 0, roiSize_norm_.width/2, 0, focal_norm_, roiSize_norm_.height/2.0f, 0, 0, 1.0f);
|
|
}
|
|
|
|
// convert vector from normalization space to camera coordinate system
|
|
cv::Mat Normalizer::cvtToCamera(cv::Point3f input, const Mat cnv_mat) {
|
|
// convert to the original camera coordinate system
|
|
Vec3f gaze_v(input.x, input.y, input.z);
|
|
// apply de-normalization
|
|
Mat gaze_v_cam = cnv_mat.inv() * Mat(gaze_v);
|
|
gaze_v_cam = gaze_v_cam / norm(gaze_v_cam);
|
|
|
|
return gaze_v_cam;
|
|
}
|
|
|
|
cv::Mat Normalizer::normalizeFace(Mat input_image, opengaze::Sample &sample) {
|
|
// get the face center in 3D space
|
|
Mat HR;
|
|
cv::Rodrigues(sample.face_patch_data.head_r, HR);
|
|
Mat HT = repeat(sample.face_patch_data.head_t, 1, 6);
|
|
Mat Fc;
|
|
add(HR*face_model_mat_, HT, Fc);
|
|
|
|
float distance = (float)norm(sample.face_patch_data.face_center); // original distance
|
|
float z_scale = distance_norm_ / distance; // scaling factor
|
|
cv::Mat scaleMat;
|
|
scaleMat = (Mat_<float>(3,3) << 1.0f, 0.0f, 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, z_scale);// scaling matrix
|
|
scaleMat.convertTo(scaleMat, CV_32F);
|
|
|
|
// get the look_at matrix
|
|
Mat hRx = HR.col(0);
|
|
Mat forward = sample.face_patch_data.face_center /distance;
|
|
Mat down = forward.cross(hRx);
|
|
down = down / norm(down);
|
|
Mat right = down.cross(forward);
|
|
right = right / norm(right);
|
|
|
|
// rotation matrix
|
|
sample.face_patch_data.face_rot = Mat(3, 3, CV_32F);
|
|
right.copyTo(sample.face_patch_data.face_rot.col(0));
|
|
down.copyTo(sample.face_patch_data.face_rot.col(1));
|
|
forward.copyTo(sample.face_patch_data.face_rot.col(2));
|
|
sample.face_patch_data.face_rot = sample.face_patch_data.face_rot.t(); // there is no scaling
|
|
sample.face_patch_data.face_rot.convertTo(sample.face_patch_data.face_rot, CV_32F);
|
|
|
|
Mat warpMat = cam_norm_ * (scaleMat * sample.face_patch_data.face_rot) * camera_matrix_.inv();// transformation matrix
|
|
// crop image and copy the equalized image
|
|
Mat face_patch;
|
|
warpPerspective(input_image, face_patch, warpMat, roiSize_norm_);
|
|
|
|
return face_patch;
|
|
}
|
|
|
|
vector<cv::Mat> Normalizer::normalizeEyes(cv::Mat input_image, Sample &sample){
|
|
vector<cv::Mat> eye_images;
|
|
|
|
Mat img_gray;
|
|
cvtColor(input_image, img_gray, CV_BGR2GRAY);
|
|
|
|
Mat eye_center;
|
|
Mat* eye_rot;
|
|
for (int i=0; i<2; ++i) {
|
|
if (i==0){
|
|
eye_center = sample.eye_data.leye_pos;
|
|
eye_rot = &sample.eye_data.leye_rot;
|
|
}
|
|
else {
|
|
eye_center = sample.eye_data.reye_pos;
|
|
eye_rot = &sample.eye_data.reye_rot;
|
|
}
|
|
|
|
|
|
float distance = (float)norm(eye_center);
|
|
float z_scale = distance_norm_ / distance;
|
|
|
|
Mat scaleMat;
|
|
scaleMat = (Mat_<float>(3,3) << 1.0f, 0.0f, 0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, z_scale);// scaling matrix
|
|
scaleMat.convertTo(scaleMat, CV_32F);
|
|
|
|
// get the look_at matrix
|
|
Mat HR;
|
|
cv::Rodrigues(sample.face_patch_data.head_r, HR);
|
|
Mat hRx = HR.col(0);
|
|
Mat forward = eye_center/distance;
|
|
Mat down = forward.cross(hRx);
|
|
down = down / norm(down);
|
|
Mat right = down.cross(forward);
|
|
right = right / norm(right);
|
|
|
|
// rotation matrix
|
|
*eye_rot = Mat(3, 3, CV_32F);
|
|
right.copyTo(eye_rot->col(0));
|
|
down.copyTo(eye_rot->col(1));
|
|
forward.copyTo(eye_rot->col(2));
|
|
*eye_rot = eye_rot->t(); // there is no scaling
|
|
|
|
Mat warpMat = cam_norm_ * (scaleMat * *eye_rot) * camera_matrix_.inv();// transformation matrix
|
|
// crop image and copy the equalized image
|
|
Mat eye_patch, eye_patch_equal;
|
|
warpPerspective(img_gray, eye_patch, warpMat, roiSize_norm_);
|
|
equalizeHist(eye_patch, eye_patch_equal);
|
|
eye_images.push_back(eye_patch_equal);
|
|
|
|
}
|
|
eye_rot = nullptr;
|
|
return eye_images;
|
|
}
|
|
|
|
void Normalizer::loadFaceModel(std::string path) {
|
|
string face_model_file_path = path + "/content/model/face_model.yml";
|
|
//
|
|
cout << endl << "Loading 3D face model for head pose estimation from : " << face_model_file_path << endl;
|
|
FileStorage fs;
|
|
if (!fs.open(face_model_file_path, FileStorage::READ)) {
|
|
cout << "Cannot load the 3D face model!" << endl;
|
|
exit(EXIT_FAILURE);
|
|
}
|
|
fs["face_model"] >> face_model_mat_;
|
|
for(int p=0; p<6; ++p)
|
|
face_model_.emplace_back(Point3d(face_model_mat_.at<float>(0,p),
|
|
face_model_mat_.at<float>(1,p),
|
|
face_model_mat_.at<float>(2,p)));
|
|
fs.release();
|
|
}
|
|
|
|
// estimate head pose via model fitting
|
|
void Normalizer::estimateHeadPose(const Point2f *landmarks, opengaze::Sample &sample) {
|
|
Mat zero_dist = Mat::zeros(1, 5, CV_64F);
|
|
vector<Point2d> landmarks_orig(landmarks,
|
|
landmarks + 6);
|
|
cv::Mat head_r, head_t;
|
|
camera_matrix_.convertTo(camera_matrix_, CV_64F); // input must be double type
|
|
solvePnP(face_model_, landmarks_orig, camera_matrix_, zero_dist, head_r, head_t, false, SOLVEPNP_DLS);
|
|
solvePnP(face_model_, landmarks_orig, camera_matrix_, zero_dist, head_r, head_t, true);
|
|
head_r.convertTo(sample.face_patch_data.head_r, CV_32F);
|
|
head_t.convertTo(sample.face_patch_data.head_t, CV_32F);
|
|
camera_matrix_.convertTo(camera_matrix_, CV_32F);
|
|
|
|
// get the face center in 3D space
|
|
Mat HR;
|
|
cv::Rodrigues(sample.face_patch_data.head_r, HR);
|
|
Mat HT = repeat(sample.face_patch_data.head_t, 1, 6);
|
|
Mat Fc;
|
|
add(HR*face_model_mat_, HT, Fc);
|
|
Mat face_center = (Fc.col(0) + Fc.col(1) + Fc.col(2) + Fc.col(3) + Fc.col(4) + Fc.col(5)) / 6.0; // face center
|
|
face_center.copyTo(sample.face_patch_data.face_center); // copy to output
|
|
sample.face_patch_data.face_center.convertTo(sample.face_patch_data.face_center, CV_32F);
|
|
|
|
Mat le = 0.5*(Fc.col(2) + Fc.col(3)); // left eye
|
|
le.copyTo(sample.eye_data.leye_pos);
|
|
sample.eye_data.leye_pos.convertTo(sample.eye_data.leye_pos, CV_32F);
|
|
Mat re = 0.5*(Fc.col(0) + Fc.col(1)); // right eye
|
|
re.copyTo(sample.eye_data.reye_pos);
|
|
sample.eye_data.reye_pos.convertTo(sample.eye_data.reye_pos, CV_32F);
|
|
|
|
}
|
|
|
|
void Normalizer::setCameraMatrix(cv::Mat input) {
|
|
camera_matrix_ = input;
|
|
camera_matrix_.convertTo(camera_matrix_, CV_32F);
|
|
}
|
|
|
|
} |