void Nieto::ExpectationMaximizationOpenCV2Features(const Mat1b &imageI, const Mat1b &imageL,
map<string, double> &i_means0, map<string, double> &i_covs0, map<string, double> &i_weights0,
map<string, double> &l_means0, map<string, double> &l_covs0, map<string, double> &l_weights0, int maxIters) {
double tempoInicio = static_cast<double>(getTickCount());
const int nFeatures = 2; // 2 features => {I, L}
const int nClusters = 4; // 4 classes => {pavement, markings, objects, unknown}
const bool aplicaResize = true;
// samples feature I
Mat1b I_imageClone = imageI.clone();
Mat1b I_trainImageClone = imageI.clone();
if (aplicaResize) resize(I_imageClone, I_trainImageClone, Size(160, 35), 0, 0, INTER_NEAREST);
Mat1d I_samples = I_trainImageClone.reshape(1, I_trainImageClone.rows * I_trainImageClone.cols);
// samples feature L
Mat1b L_imageClone = imageL.clone();
Mat1b L_trainImageClone = imageL.clone();
if (aplicaResize) resize(L_imageClone, L_trainImageClone, Size(160, 35), 0, 0, INTER_NEAREST);
Mat1d L_samples = L_trainImageClone.reshape(1, L_trainImageClone.rows * L_trainImageClone.cols);
// junta as amostras (uma em cada linha)
Mat1d samplesArray[] = { I_samples, L_samples };
Mat1d samples;
cv::hconcat(samplesArray, 2, samples);
// formata o _means0
Mat1d means0 = Mat1d(nClusters, nFeatures, CV_64FC1);
means0.at<double>(0, 0) = i_means0["pavement"];
means0.at<double>(1, 0) = i_means0["markings"];
means0.at<double>(2, 0) = i_means0["objects"];
means0.at<double>(3, 0) = 255.0 / 2.0;
means0.at<double>(0, 1) = l_means0["pavement"];
means0.at<double>(1, 1) = l_means0["markings"];
means0.at<double>(2, 1) = l_means0["objects"];
means0.at<double>(3, 1) = 255.0 / 2.0;
// formata o _covs0
Mat1d covs0_pavement = Mat1d(Size(nFeatures, nFeatures), double(0));
covs0_pavement.at<double>(0, 0) = i_covs0["pavement"];
covs0_pavement.at<double>(1, 1) = l_covs0["pavement"];
Mat1d covs0_markings = Mat1d(Size(nFeatures, nFeatures), double(0));;
covs0_markings.at<double>(0, 0) = i_covs0["markings"];
covs0_markings.at<double>(1, 1) = l_covs0["markings"];
Mat1d covs0_objects = Mat1d(Size(nFeatures, nFeatures), double(0));;
covs0_objects.at<double>(0, 0) = i_covs0["objects"];
covs0_objects.at<double>(1, 1) = l_covs0["objects"];
Mat1d covs0_unknown = Mat1d(Size(nFeatures, nFeatures), double(0));;
covs0_unknown.at<double>(0, 0) = ((255.0 / 2.0) / sqrt(3)) * ((255.0 / 2.0) / sqrt(3));
covs0_unknown.at<double>(1, 1) = ((255.0 / 2.0) / sqrt(3)) * ((255.0 / 2.0) / sqrt(3));
vector<Mat> covs0 = {
covs0_pavement,
covs0_markings,
covs0_objects,
covs0_unknown
};
// formata o _weights0
Mat1d weights0 = Mat1d(nClusters, 1, CV_64FC1);
double total_i = i_weights0["pavement"] + i_weights0["markings"] + i_weights0["objects"] + i_weights0["unknown"];
double total_l = l_weights0["pavement"] + l_weights0["markings"] + l_weights0["objects"] + l_weights0["unknown"];
double total_weights = total_i + total_l;
weights0.at<double>(0, 0) = (i_weights0["pavement"] + l_weights0["pavement"]) / total_weights;
weights0.at<double>(1, 0) = (i_weights0["markings"] + l_weights0["markings"]) / total_weights;
weights0.at<double>(2, 0) = (i_weights0["objects"] + l_weights0["objects"]) / total_weights;
weights0.at<double>(3, 0) = (i_weights0["unknown"] + l_weights0["unknown"]) / total_weights;
// cout << means0 << endl;
// condi��es do EM
// dims => samples.cols
// if (!(&means0) || (!means0.empty() && means0.rows == nClusters && means0.cols == samples.cols && means0.channels() == 1)) cout << "means - ok!" << endl;
EM em = EM(nClusters, EM::COV_MAT_DIAGONAL);
em.set("maxIters", maxIters);
em.trainE(samples, means0, covs0, weights0);
// calcula o tempo de execu��o
double tempoFim = static_cast<double>(getTickCount());
double tempoExecutando = ((tempoFim - tempoInicio) / getTickFrequency()) * 1000;
// exibe as sa�das definidas (texto e/ou imagem)
if (verbose) cout << "- em opencv (2 features): " << tempoExecutando << " ms" << endl;
if (display) {
// predict
Mat1b predictedImage = Mat1b(I_imageClone.size(), uchar(0));
for (int j = 0; j < predictedImage.rows; ++j) {
unsigned char *ptRowI = I_imageClone.ptr<uchar>(j);
unsigned char *ptRowL = L_imageClone.ptr<uchar>(j);
unsigned char *ptRowDst = predictedImage.ptr<uchar>(j);
for (int i = 0; i < predictedImage.cols; ++i) {
Mat1d elementPredict = Mat1d(Size(2, 1), CV_64FC1);
elementPredict.at<double>(0) = ptRowL[i];
elementPredict.at<double>(1) = ptRowI[i];
Vec2d emPredicted = em.predict(elementPredict);
switch ((int)emPredicted[1]) {
case 0: ptRowDst[i] = 160; break;
case 1: ptRowDst[i] = 255; break;
case 2: ptRowDst[i] = 80; break;
case 3: ptRowDst[i] = 0; break;
}
}
}
imshow("EM OpenCV - 2 Features", predictedImage);
}
}
void Nieto::ExpectationMaximizationOpenCV(const Mat1b &inGrayFrameRoi, int maxIters, map<string, double> &_means0, map<string, double> &_covs0, map<string, double> &_weights0) {
double tempoInicio = static_cast<double>(getTickCount());
const int nClusters = 4; // 4 classes => {pavement, markings, objects, unknown}
const bool aplicaResize = true;
EM em = EM(nClusters, EM::COV_MAT_DIAGONAL);
Mat1b grayFrameRoiClone = inGrayFrameRoi.clone();
Mat1b trainGrayFrameRoiClone = inGrayFrameRoi.clone();
if (aplicaResize) resize(grayFrameRoiClone, trainGrayFrameRoiClone, Size(160, 35), 0, 0, INTER_NEAREST);
Mat1d samples = trainGrayFrameRoiClone.reshape(1, trainGrayFrameRoiClone.rows * trainGrayFrameRoiClone.cols);
// formata o _means0
Mat1d means0 = Mat1d(nClusters, 1, CV_64FC1);
means0.at<double>(0) = _means0["pavement"];
means0.at<double>(1) = _means0["markings"];
means0.at<double>(2) = _means0["objects"];
means0.at<double>(3) = 255.0 / 2.0;
// formata o _covs0
vector<Mat> covs0 = {
Mat1d(1, 1, _covs0["pavement"]),
Mat1d(1, 1, _covs0["markings"]),
Mat1d(1, 1, _covs0["objects"]),
Mat1d(1, 1, ((255.0 / 2.0) / sqrt(3)) * ((255.0 / 2.0) / sqrt(3)))
};
// formata o _weights0
// Mat1d weights0 = *(Mat1f(nClusters, 1, CV_64FC1) << 0.75, 0.10, 0.10, 0.05);
Mat1d weights0 = *(Mat1f(nClusters, 1, CV_64FC1) <<
_weights0["pavement"],
_weights0["markings"],
_weights0["objects"],
_weights0["unknown"]
);
// cout << means0 << endl;
em.set("maxIters", maxIters);
em.trainE(samples, means0, covs0, weights0);
// calcula o tempo de execu��o
double tempoFim = static_cast<double>(getTickCount());
double tempoExecutando = ((tempoFim - tempoInicio) / getTickFrequency()) * 1000;
// exibe as sa�das definidas (texto e/ou imagem)
if (verbose) cout << "- em opencv (1 feature): " << tempoExecutando << " ms" << endl;
if (display) {
// predict
Mat1b predictedImage = Mat1b(grayFrameRoiClone.size(), uchar(0));
for (int j = 0; j < predictedImage.rows; ++j) {
unsigned char *ptRowSrc = grayFrameRoiClone.ptr<uchar>(j);
unsigned char *ptRowDst = predictedImage.ptr<uchar>(j);
for (int i = 0; i < predictedImage.cols; ++i) {
Vec2d emPredicted = em.predict(ptRowSrc[i]);
switch ((int)emPredicted[1]) {
case 0: ptRowDst[i] = 160; break;
case 1: ptRowDst[i] = 255; break;
case 2: ptRowDst[i] = 80; break;
case 3: ptRowDst[i] = 0; break;
}
}
}
imshow("EM OpenCV - 1 Feature", predictedImage);
}
}
