tensor2D eigenVectors(const tensor2D& t)
{
vector2D evals(eigenValues(t));
tensor2D evs
(
eigenVector(t, evals.x()),
eigenVector(t, evals.y())
);
return evs;
}
tensor eigenVectors(const symmTensor& t)
{
vector evals(eigenValues(t));
tensor evs
(
eigenVector(t, evals.x()),
eigenVector(t, evals.y()),
eigenVector(t, evals.z())
);
return evs;
}
TqrEigenDecomposition::TqrEigenDecomposition(const Array& diag,
const Array& sub,
EigenVectorCalculation calc,
ShiftStrategy strategy)
: iter_(0), d_(diag),
ev_((calc == WithEigenVector)? d_.size() :
(calc == WithoutEigenVector)? 0 : 1, d_.size(), 0)
{
Size n = diag.size();
QL_REQUIRE(n == sub.size()+1, "Wrong dimensions");
Array e(n, 0.0);
std::copy(sub.begin(),sub.end(),e.begin()+1);
Size i;
for (i=0; i < ev_.rows(); ++i) {
ev_[i][i] = 1.0;
}
for (Size k=n-1; k >=1; --k) {
while (!offDiagIsZero(k, e)) {
Size l = k;
while (--l > 0 && !offDiagIsZero(l,e));
iter_++;
Real q = d_[l];
if (strategy != NoShift) {
// calculated eigenvalue of 2x2 sub matrix of
// [ d_[k-1] e_[k] ]
// [ e_[k] d_[k] ]
// which is closer to d_[k+1].
// FLOATING_POINT_EXCEPTION
const Real t1 = std::sqrt(
0.25*(d_[k]*d_[k] + d_[k-1]*d_[k-1])
- 0.5*d_[k-1]*d_[k] + e[k]*e[k]);
const Real t2 = 0.5*(d_[k]+d_[k-1]);
const Real lambda =
(std::fabs(t2+t1 - d_[k]) < std::fabs(t2-t1 - d_[k]))?
t2+t1 : t2-t1;
if (strategy == CloseEigenValue) {
q-=lambda;
} else {
q-=((k==n-1)? 1.25 : 1.0)*lambda;
}
}
// the QR transformation
Real sine = 1.0;
Real cosine = 1.0;
Real u = 0.0;
bool recoverUnderflow = false;
for (Size i=l+1; i <= k && !recoverUnderflow; ++i) {
const Real h = cosine*e[i];
const Real p = sine*e[i];
e[i-1] = std::sqrt(p*p+q*q);
if (e[i-1] != 0.0) {
sine = p/e[i-1];
cosine = q/e[i-1];
const Real g = d_[i-1]-u;
const Real t = (d_[i]-g)*sine+2*cosine*h;
u = sine*t;
d_[i-1] = g + u;
q = cosine*t - h;
for (Size j=0; j < ev_.rows(); ++j) {
const Real tmp = ev_[j][i-1];
ev_[j][i-1] = sine*ev_[j][i] + cosine*tmp;
ev_[j][i] = cosine*ev_[j][i] - sine*tmp;
}
} else {
// recover from underflow
d_[i-1] -= u;
e[l] = 0.0;
recoverUnderflow = true;
}
}
if (!recoverUnderflow) {
d_[k] -= u;
e[k] = q;
e[l] = 0.0;
}
}
}
// sort (eigenvalues, eigenvectors),
// code taken from symmetricSchureDecomposition.cpp
std::vector<std::pair<Real, std::vector<Real> > > temp(n);
std::vector<Real> eigenVector(ev_.rows());
for (i=0; i<n; i++) {
if (ev_.rows() > 0)
std::copy(ev_.column_begin(i),
ev_.column_end(i), eigenVector.begin());
temp[i] = std::make_pair(d_[i], eigenVector);
}
std::sort(temp.begin(), temp.end(),
std::greater<std::pair<Real, std::vector<Real> > >());
// first element is positive
for (i=0; i<n; i++) {
d_[i] = temp[i].first;
Real sign = 1.0;
if (ev_.rows() > 0 && temp[i].second[0]<0.0)
sign = -1.0;
for (Size j=0; j<ev_.rows(); ++j) {
ev_[j][i] = sign * temp[i].second[j];
}
}
}
void CVDataFilter::Paint(CVPipeline * pipe)
{
if (returnType == CVReturnType::QUADS)
{
// Draw quads.
if (!pipe->quads.Size())
return;
// Copy original input
pipe->initialInput.copyTo(pipe->output);
// Convert to color if needed.
int channelsBefore = pipe->output.channels();
if (channelsBefore == 1)
{
// pipe->output.convertTo(pipe->output, CV_8UC3);
cv::cvtColor(pipe->output, pipe->output, CV_GRAY2RGB);
}
int channelsAfter = pipe->output.channels();
// o.o Paste!
for (int i = 0; i < pipe->quads.Size(); ++i)
{
Quad quad = pipe->quads[i];
#define RGB(r,g,b) cv::Scalar(b,g,r)
rng = cv::RNG(1344);
cv::Scalar color = RGB(rng.uniform(0, 255),rng.uniform(0, 255),rng.uniform(0, 255));
// cv::Mat rectImage = cv::Mat::zeros(pipe->output.size(), CV_8UC3);
cv::rectangle(pipe->output, cv::Point(quad.point1.x, quad.point1.y), cv::Point(quad.point3.x, quad.point3.y), color, CV_FILLED);
float alpha = 0.2f;
float beta = 1 - alpha;
// cv::addWeighted(rectImage, alpha, pipe->output, beta, 0.0, pipe->output);
// rectImage.copyTo(pipe->output);
}
}
else if (returnType == CVReturnType::APPROXIMATED_POLYGONS)
{
// Copy original input..
pipe->initialInput.copyTo(pipe->output);
RenderPolygons(pipe);
}
else if (returnType == CVReturnType::CV_IMAGE)
{
// Do nothing as the image should already be in the ouput matrix of the pipeline.
}
else if (returnType == CVReturnType::LINES ||
returnType == CVReturnType::CV_LINES)
{
// Convert the color to colors again for visualization...
pipe->initialInput.copyTo(pipe->output);
for( size_t i = 0; i < pipe->lines.Size(); i++ )
{
Line & line = pipe->lines[i];
// Multiply the line coordinates with the last used inverted scale.
line.start *= pipe->currentScaleInv;
line.stop *= pipe->currentScaleInv;
cv::line( pipe->output, cv::Point(line.start.x, line.start.y),
cv::Point(line.stop.x, line.stop.y), cv::Scalar(0,0,255), 3, 8 );
}
}
else if (returnType == CVReturnType::CV_CONTOURS)
{
pipe->initialInput.copyTo(pipe->output);
RenderContours(pipe);
}
switch(returnType)
{
case CVReturnType::CV_CONTOUR_SEGMENTS:
{
// Render shit!
pipe->initialInput.copyTo(pipe->output);
RenderContourSegments(pipe);
break;
}
}
if (returnType == CVReturnType::CV_CONTOUR_ELLIPSES)
{
pipe->initialInput.copyTo(pipe->output);
RenderContours(pipe);
RenderContourBounds(pipe);
}
else if (returnType == CVReturnType::CV_CONVEX_HULLS)
{
pipe->initialInput.copyTo(pipe->output);
RenderContours(pipe);
RenderConvexHulls(pipe);
}
else if (returnType == CVReturnType::CV_CONVEXITY_DEFECTS)
{
pipe->initialInput.copyTo(pipe->output);
RenderContours(pipe);
RenderConvexHulls(pipe);
RenderConvexityDefects(pipe);
}
else if (returnType == CVReturnType::HANDS)
{
// Convert image to RGB for easier display
int channelsBefore = pipe->initialInput.channels();
// cv::cvtColor(*pipe->initialInput, pipe->output, CV_GRAY2RGB);
pipe->initialInput.copyTo(pipe->output);
int channelsAfter = pipe->output.channels();
// Check if we got contour-segments, if so prioritize them.
RenderHands(pipe);
}
else if (returnType == CVReturnType::FINGER_STATES)
{
// Render the last known one.
pipe->initialInput.copyTo(pipe->output);
FingerState & state = pipe->fingerStates.Last();
cv::Scalar color(255,0,0,255);
for (int i = 0; i < state.positions.Size(); ++i)
{
Vector3f pos = state.positions[i];
cv::circle(pipe->output, cv::Point(pos.x, pos.y), 5, color, 3);
}
}
else if (returnType == CVReturnType::POINT_CLOUDS)
{
pipe->initialInput.copyTo(pipe->output);
for (int i = 0; i < pipe->pointClouds.Size(); ++i)
{
int r,g,b,a;
r = g = b = a = 255;
if (i % 2 == 0)
g = 0;
if (i % 4 == 0)
b = 0;
cv::Scalar color(r,g,b,a);
CVPointCloud & pc = pipe->pointClouds[i];
for (int j = 0; j < pc.points.Size(); ++j)
{
Vector2f & pos = pc.points[j];
cv::circle(pipe->output, cv::Point(pos.x, pos.y), 2, color, 3);
}
/// Render PCA data if applicable.
for (int j = 0; j < pc.eigenVectors.Size(); ++j)
{
cv::Scalar color = j == 0? CV_RGB(255, 255, 0) : CV_RGB(0, 255, 255);
cv::Point2f eigenVector(pc.eigenVectors[j].x, pc.eigenVectors[j].y);
float eigenValue = pc.eigenValues[j];
cv::Point2f scaledEigenVector = eigenVector * eigenValue * 0.02;
cv::Point2f pcaCenter(pc.pcaCenter.x, pc.pcaCenter.y);
circle(pipe->output, pcaCenter, 3, CV_RGB(255, 0, 255), 2);
line(pipe->output, pcaCenter, pcaCenter + scaledEigenVector, color);
}
}
}
else if (returnType == CVReturnType::CV_OPTICAL_FLOW)
{
// http://stackoverflow.com/questions/7693561/opencv-displaying-a-2-channel-image-optical-flow
// Split the coordinates.
/*
cv::Mat xy[2];
cv::split(pipe->opticalFlow, xy);
cv::Mat magnitude, angle;
// Fetch matrix from pipeline.
cv::cartToPolar(xy[0], xy[1], magnitude, angle, true);
double magMax;
cv::minMaxLoc(magnitude, 0, &magMax);
magnitude.convertTo(magnitude, -1, 1.0 / magMax);
// Build HSV image?
cv::Mat hsvChannels[3], hsv;
hsvChannels[0] = angle;
hsvChannels[1] = cv::Mat::ones(angle.size(), CV_32F);
hsvChannels[2] = magnitude;
cv::merge(hsvChannels, 3, hsv);
// Convert to rgb and send to output.
cv::cvtColor(hsv, pipe->output, cv::COLOR_HSV2RGB);
*/
}
else if (returnType == CVReturnType::NOTHING)
{
// Nothing to render...
}
else if (returnType == -1)
{
// Nothing to render if error.
std::cout<<"\nreturnType variable in filter "<<name<<" not set!";
}
else if (returnType == CVReturnType::RENDER)
// Nothing to render if render.
;
else
; // std::cout<<"\nCVDataFilter::Paint called. Forgot to subclass the paint-method?";
}
PCA&
PCA::fit( const std::vector< std::vector< double > >& data )
{
const size_t nSamples = data.size();
const size_t nFeatures = data.front().size();
// Calculate the means vector
arma::mat meansVector( nFeatures, 1 );
for ( size_t iFeature = 0; iFeature < nFeatures; ++iFeature ) {
double sx = 0;
for ( size_t iSample = 0; iSample < nSamples; ++iSample ) {
const double value = data[iSample][iFeature];
if ( std::isnan(value) )
throw std::runtime_error( "PCA::fit : nan value encountered at input!" );
sx += value;
}
meansVector(iFeature, 0) = sx / nSamples;
}
// Construct the covariance matrix from the scatter matrix
arma::mat covMatrix( nFeatures, nFeatures, arma::fill::zeros );
for ( size_t iSample = 0; iSample < nSamples; ++iSample ) {
arma::mat sampleData( data[iSample ] );
arma::mat d = sampleData - meansVector;
covMatrix += d * d.t();
}
covMatrix /= nSamples;
// Now find the eigenvalues and eigenvectors
arma::cx_vec eigval;
arma::cx_mat eigvec;
bool result = arma::eig_gen(eigval, eigvec, covMatrix );
if (! result ) {
throw std::runtime_error("PCA::fit : eigenvalue decomposition failed!");
}
// Normalise the eigenvalues
m_eigPairs.clear();
m_eigPairs.reserve( nFeatures );
double eigSum = 0;
for ( size_t iFeature = 0; iFeature < nFeatures; ++iFeature ) {
const double eigMagnitude = std::abs( eigval(iFeature) );
arma::cx_vec eigenVectorFromCalculation = eigvec.row( iFeature );
std::vector<double> eigenVector( nFeatures, 0.0 );
for (size_t j = 0; j < nFeatures; ++j ) eigenVector[j] = eigenVectorFromCalculation(j).real();
m_eigPairs.push_back( std::make_pair( eigMagnitude,
eigenVector ) );
eigSum += eigMagnitude;
}
for ( size_t iFeature = 0; iFeature < nFeatures; ++iFeature ) {
m_eigPairs[iFeature].first /= eigSum;
}
// Sort the eigenValues
std::sort( m_eigPairs.begin(), m_eigPairs.end(),
[] (const std::pair<double, std::vector<double> >&a,
const std::pair<double, std::vector<double> >&b ) { return a.first > b.first; } );
return *this;
}
SymmetricSchurDecomposition::SymmetricSchurDecomposition(const Matrix & s)
: diagonal_(s.rows()), eigenVectors_(s.rows(), s.columns(), 0.0) {
QL_REQUIRE(s.rows() > 0 && s.columns() > 0, "null matrix given");
QL_REQUIRE(s.rows()==s.columns(), "input matrix must be square");
Size size = s.rows();
for (Size q=0; q<size; q++) {
diagonal_[q] = s[q][q];
eigenVectors_[q][q] = 1.0;
}
Matrix ss = s;
std::vector<Real> tmpDiag(diagonal_.begin(), diagonal_.end());
std::vector<Real> tmpAccumulate(size, 0.0);
Real threshold, epsPrec = 1e-15;
bool keeplooping = true;
Size maxIterations = 100, ite = 1;
do {
//main loop
Real sum = 0;
for (Size a=0; a<size-1; a++) {
for (Size b=a+1; b<size; b++) {
sum += std::fabs(ss[a][b]);
}
}
if (sum==0) {
keeplooping = false;
} else {
/* To speed up computation a threshold is introduced to
make sure it is worthy to perform the Jacobi rotation
*/
if (ite<5) threshold = 0.2*sum/(size*size);
else threshold = 0.0;
Size j, k, l;
for (j=0; j<size-1; j++) {
for (k=j+1; k<size; k++) {
Real sine, rho, cosin, heig, tang, beta;
Real smll = std::fabs(ss[j][k]);
if(ite> 5 &&
smll<epsPrec*std::fabs(diagonal_[j]) &&
smll<epsPrec*std::fabs(diagonal_[k])) {
ss[j][k] = 0;
} else if (std::fabs(ss[j][k])>threshold) {
heig = diagonal_[k]-diagonal_[j];
if (smll<epsPrec*std::fabs(heig)) {
tang = ss[j][k]/heig;
} else {
beta = 0.5*heig/ss[j][k];
tang = 1.0/(std::fabs(beta)+
std::sqrt(1+beta*beta));
if (beta<0)
tang = -tang;
}
cosin = 1/std::sqrt(1+tang*tang);
sine = tang*cosin;
rho = sine/(1+cosin);
heig = tang*ss[j][k];
tmpAccumulate[j] -= heig;
tmpAccumulate[k] += heig;
diagonal_[j] -= heig;
diagonal_[k] += heig;
ss[j][k] = 0.0;
for (l=0; l+1<=j; l++)
jacobiRotate_(ss, rho, sine, l, j, l, k);
for (l=j+1; l<=k-1; l++)
jacobiRotate_(ss, rho, sine, j, l, l, k);
for (l=k+1; l<size; l++)
jacobiRotate_(ss, rho, sine, j, l, k, l);
for (l=0; l<size; l++)
jacobiRotate_(eigenVectors_,
rho, sine, l, j, l, k);
}
}
}
for (k=0; k<size; k++) {
tmpDiag[k] += tmpAccumulate[k];
diagonal_[k] = tmpDiag[k];
tmpAccumulate[k] = 0.0;
}
}
} while (++ite<=maxIterations && keeplooping);
QL_ENSURE(ite<=maxIterations,
"Too many iterations (" << maxIterations << ") reached");
// sort (eigenvalues, eigenvectors)
std::vector<std::pair<Real, std::vector<Real> > > temp(size);
std::vector<Real> eigenVector(size);
Size row, col;
for (col=0; col<size; col++) {
std::copy(eigenVectors_.column_begin(col),
eigenVectors_.column_end(col), eigenVector.begin());
temp[col] = std::make_pair(diagonal_[col], eigenVector);
}
std::sort(temp.begin(), temp.end(),
std::greater<std::pair<Real, std::vector<Real> > >());
Real maxEv = temp[0].first;
for (col=0; col<size; col++) {
// check for round-off errors
diagonal_[col] =
(std::fabs(temp[col].first/maxEv)<1e-16 ? 0.0 :
temp[col].first);
Real sign = 1.0;
if (temp[col].second[0]<0.0)
sign = -1.0;
for (row=0; row<size; row++) {
eigenVectors_[row][col] = sign * temp[col].second[row];
}
}
}
