RcppExport SEXP rthxpos(SEXP m)
{
Rcpp::NumericMatrix tmpm = Rcpp::NumericMatrix(m);
int nr = tmpm.nrow();
int nc = tmpm.ncol();
thrust::device_vector<double> dmat(tmpm.begin(),tmpm.end());
// make space for the transpose
thrust::device_vector<double> dxp(nr*nc);
// iterator to march through the matrix elements
thrust::counting_iterator<int> seqb(0);
thrust::counting_iterator<int> seqe = seqb + nr*nc;
// for each i in seq, copy the matrix elt to its spot in the
// transpose
thrust::for_each(seqb,seqe,
copyelt2xp(dmat.begin(),dxp.begin(),nr,nc));
// prepare the R output, and return it
Rcpp::NumericVector routmat(nc*nr);
thrust::copy(dxp.begin(),dxp.end(),routmat.begin());
return routmat;
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
//L2L2OpticalFlow(u1,u2,tol,lambda,ux,uy,ut)
{
double *u1 = mxGetPr(prhs[0]);
double *u2 = mxGetPr(prhs[1]);
float tol = (float)mxGetScalar(prhs[2]);
float lambda = (float)mxGetScalar(prhs[3]);
int maxIterations = (int)mxGetScalar(prhs[4]);
const size_t *sizeImage = mxGetDimensions(prhs[0]);
double *inputV = 0;
double *inputY = 0;
int typeNorm = 1;
if (nrhs > 5)
{
typeNorm = (int)mxGetScalar(prhs[5]);
}
if (nrhs > 6)
{
inputV = mxGetPr(prhs[6]);
}
if (nrhs > 7)
{
inputY = mxGetPr(prhs[7]);
}
int nPx = (int)(sizeImage[0] * sizeImage[1]);
const size_t sizeY[2] = {5*nPx,1};
// Output v1
plhs[0] = mxCreateNumericArray(2, sizeImage, mxDOUBLE_CLASS, mxREAL);
double *Outv1 = mxGetPr(plhs[0]);
// Output v2
plhs[1] = mxCreateNumericArray(2, sizeImage, mxDOUBLE_CLASS, mxREAL);
double *Outv2 = mxGetPr(plhs[1]);
// Output Y
plhs[2] = mxCreateNumericArray(2, sizeY, mxDOUBLE_CLASS, mxREAL);
double *YOut = mxGetPr(plhs[2]);
float* v1 = new float[nPx];
float* v2 = new float[nPx];
float* u = new float[nPx];
float* ut = new float[nPx];
float* sigut = new float[nPx];
float* y11 = new float[nPx];
float* y12 = new float[nPx];
float* y21 = new float[nPx];
float* y22 = new float[nPx];
float* y5 = new float[nPx];
float* Kty1 = new float[nPx];
float* Kty2 = new float[nPx];
float* Kx11 = new float[nPx];
float* Kx12 = new float[nPx];
float* Kx21 = new float[nPx];
float* Kx22 = new float[nPx];
float* Kx5 = new float[nPx];
float tau = 1.0f / sqrt(8.0f);
float sigma = tau;
float dTau = 1.0f / tau;
float dSigma = 1.0f / sigma;
//residuals
float p = 0;
float d = 0;
float err = 1.0;
float ssl = 1.0f - sigma / (sigma + lambda);
#pragma omp parallel for
for (int j = 0; j < sizeImage[1]; ++j)
{
for (int i = 0; i < sizeImage[0]; ++i)
{
int tmpIndex = index2DtoLinear(sizeImage, i, j);
//Index for gradients
u[tmpIndex] = (float)u1[tmpIndex];
ut[tmpIndex] = (float)(u2[tmpIndex] - u1[tmpIndex]);
sigut[tmpIndex] = (float)(sigma*ut[tmpIndex]);
if (nrhs > 6)
{
v1[tmpIndex] = (float)inputV[tmpIndex];
v2[tmpIndex] = (float)inputV[nPx + tmpIndex];
}
else
{
v1[tmpIndex] = 0;
v2[tmpIndex] = 0;
}
Kty1[tmpIndex] = 0;
Kty2[tmpIndex] = 0;
if (nrhs > 7)
{
y11[tmpIndex] = (float)inputY[tmpIndex];
y12[tmpIndex] = (float)inputY[nPx + tmpIndex];
y21[tmpIndex] = (float)inputY[2 * nPx + tmpIndex];
y22[tmpIndex] = (float)inputY[3 * nPx + tmpIndex];
y5[tmpIndex] = (float)inputY[4 * nPx + tmpIndex];
}
else
{
y11[tmpIndex] = 0;
y12[tmpIndex] = 0;
y21[tmpIndex] = 0;
y22[tmpIndex] = 0;
y5[tmpIndex] = 0;
}
Kx11[tmpIndex] = 0;
Kx12[tmpIndex] = 0;
Kx21[tmpIndex] = 0;
Kx22[tmpIndex] = 0;
Kx5[tmpIndex] = 0;
}
}
int iterations = 0;
while (err > tol && iterations <= maxIterations)
{
++iterations;
if (iterations % 50 == 0)
{
p = 0;
d = 0;
}
//primal step
#pragma omp parallel for reduction(+:p)
for (int j = 0; j < sizeImage[1]; ++j)
{
for (int i = 0; i < sizeImage[0]; ++i)
{
int tmpIndex = index2DtoLinear(sizeImage, i, j);
float Kty1Old = Kty1[tmpIndex];
float Kty2Old = Kty2[tmpIndex];
//transpose equals -div
Kty1[tmpIndex] = -(dxm(y11, sizeImage, i, j) + dym(y12, sizeImage, i, j) + dycT(y5, u, sizeImage, i, j));
Kty2[tmpIndex] = -(dxm(y21, sizeImage, i, j) + dym(y22, sizeImage, i, j) + dxcT(y5, u, sizeImage, i, j));
float v1Old = v1[tmpIndex];
float v2Old = v2[tmpIndex];
v1[tmpIndex] = v1Old - tau*Kty1[tmpIndex];
v2[tmpIndex] = v2Old - tau*Kty2[tmpIndex];
if (iterations % 50 == 0)
{
//residuals
p += myAbs((v1Old - v1[tmpIndex]) * dTau - Kty1Old + Kty1[tmpIndex])
+ myAbs((v2Old - v2[tmpIndex]) * dTau - Kty2Old + Kty2[tmpIndex]);
}
}
}
//dual step
#pragma omp parallel for reduction(+:d)
for (int j = 0; j < sizeImage[1]; ++j)
{
for (int i = 0; i < sizeImage[0]; ++i)
{
int tmpIndex = index2DtoLinear(sizeImage, i, j);
float Kx11Old = Kx11[tmpIndex];
float Kx12Old = Kx12[tmpIndex];
float Kx21Old = Kx21[tmpIndex];
float Kx22Old = Kx22[tmpIndex];
float Kx5Old = Kx5[tmpIndex];
Kx11[tmpIndex] = dxp(v1, sizeImage, i, j);
Kx12[tmpIndex] = dyp(v1, sizeImage, i, j);
Kx21[tmpIndex] = dxp(v2, sizeImage, i, j);
Kx22[tmpIndex] = dyp(v2, sizeImage, i, j);
Kx5[tmpIndex] = dyc(v1, u, sizeImage, i, j) + dxc(v2, u, sizeImage, i, j);
float y11Old = y11[tmpIndex];
float y12Old = y12[tmpIndex];
float y21Old = y21[tmpIndex];
float y22Old = y22[tmpIndex];
float y5Old = y5[tmpIndex];
y11[tmpIndex] = ssl * (y11[tmpIndex] + sigma*(2 * Kx11[tmpIndex] - Kx11Old));
y12[tmpIndex] = ssl * (y12[tmpIndex] + sigma*(2 * Kx12[tmpIndex] - Kx12Old));
y21[tmpIndex] = ssl * (y21[tmpIndex] + sigma*(2 * Kx21[tmpIndex] - Kx21Old));
y22[tmpIndex] = ssl * (y22[tmpIndex] + sigma*(2 * Kx22[tmpIndex] - Kx22Old));
y5[tmpIndex] = myMax(-1.0f, myMin(1.0f, y5[tmpIndex] + sigma*(2 * Kx5[tmpIndex] - Kx5Old) + sigut[tmpIndex]));
if (iterations % 50 == 0)
{
d += myAbs((y11Old - y11[tmpIndex]) * dSigma - Kx11Old + Kx11[tmpIndex]) +
myAbs((y12Old - y12[tmpIndex]) * dSigma - Kx12Old + Kx12[tmpIndex]) +
myAbs((y21Old - y21[tmpIndex]) * dSigma - Kx21Old + Kx21[tmpIndex]) +
myAbs((y22Old - y22[tmpIndex]) * dSigma - Kx22Old + Kx22[tmpIndex]) +
myAbs((y5Old - y5[tmpIndex]) * dSigma - Kx5Old + Kx5[tmpIndex]);
}
}
}
if (iterations % 50 == 0)
{
err = (d*d + p*p) / nPx;
}
if (iterations % 1000 == 0)
{
mexPrintf("Iteration %d,Residual %e\n", iterations, err);
mexEvalString("drawnow;");
}
}
//write output
#pragma omp parallel for
for (int j = 0; j < sizeImage[1]; ++j)
{
for (int i = 0; i < sizeImage[0]; ++i)
{
int tmpIndex = index2DtoLinear(sizeImage, i, j);
YOut[tmpIndex] = (double)y11[tmpIndex];
YOut[tmpIndex + nPx] = (double)y12[tmpIndex];
YOut[tmpIndex + 2 * nPx] = (double)y21[tmpIndex];
YOut[tmpIndex + 3 * nPx] = (double)y22[tmpIndex];
YOut[tmpIndex + 4 * nPx] = (double)y5[tmpIndex];
Outv1[tmpIndex] = (double) v1[tmpIndex];
Outv2[tmpIndex] = (double) v2[tmpIndex];
}
}
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
//L2L2OpticalFlow(u1,u2,tol,lambda,ux,uy,ut)
{
double *u1 = mxGetPr(prhs[0]);
double *u2 = mxGetPr(prhs[1]);
float tol = (float)mxGetScalar(prhs[2]);
float lambda = (float)mxGetScalar(prhs[3]);
int maxIterations = (int)mxGetScalar(prhs[4]);
const mwSize *sizeImage = mxGetDimensions(prhs[0]);
double *inputV = 0;
double *inputY = 0;
int typeNorm = 1;
if (nrhs > 5)
{
typeNorm = (int)mxGetScalar(prhs[5]);
}
if (nrhs > 6)
{
inputV = mxGetPr(prhs[6]);
}
if (nrhs > 7)
{
inputY = mxGetPr(prhs[7]);
}
int nPx = (int)(sizeImage[0] * sizeImage[1]);
const mwSize sizeY[2] = {4*nPx,1};
// Output v1
plhs[0] = mxCreateNumericArray(2, sizeImage, mxDOUBLE_CLASS, mxREAL);
double *Outv1 = mxGetPr(plhs[0]);
// Output v2
plhs[1] = mxCreateNumericArray(2, sizeImage, mxDOUBLE_CLASS, mxREAL);
double *Outv2 = mxGetPr(plhs[1]);
// Output Y
plhs[2] = mxCreateNumericArray(2, sizeY, mxDOUBLE_CLASS, mxREAL);
double *YOut = mxGetPr(plhs[2]);
float* v1 = new float[nPx];
float* v2 = new float[nPx];
float* ux = new float[nPx];
float* uy = new float[nPx];
float* ut = new float[nPx];
float* uxut = new float[nPx];
float* uyut = new float[nPx];
float* c1 = new float[nPx];
float* c2 = new float[nPx];
float* c3 = new float[nPx];
float* teiler = new float[nPx];
float* y11 = new float[nPx];
float* y12 = new float[nPx];
float* y21 = new float[nPx];
float* y22 = new float[nPx];
float* Kty1 = new float[nPx];
float* Kty2 = new float[nPx];
float* Kx11 = new float[nPx];
float* Kx12 = new float[nPx];
float* Kx21 = new float[nPx];
float* Kx22 = new float[nPx];
float tau = 1.0f / sqrt(8.0f);
float sigma = tau;
float dTau = 1.0f / tau;
float dSigma = 1.0f / sigma;
//residuals
float p = 0;
float d = 0;
float err = 1.0;
float ssl = 1.0f - sigma / (sigma + lambda);
int i, j;
#pragma omp parallel for private(i,j)
for (j = 0; j < sizeImage[1]; ++j)
{
for (i = 0; i < sizeImage[0]; ++i)
{
int tmpIndex = index2DtoLinear(sizeImage, i, j);
//Index for gradients
ut[tmpIndex] = (float)(u2[tmpIndex] - u1[tmpIndex]);
if (i>0 && i < sizeImage[0] - 1)
{
uy[tmpIndex] = (float)(0.5f * (u1[index2DtoLinear(sizeImage, i + 1, j)] - u1[index2DtoLinear(sizeImage, i - 1, j)]));
}
else
{
uy[tmpIndex] = 0.0f;
}
if (j>0 && j < sizeImage[1] - 1)
{
ux[tmpIndex] = (float)(0.5f * (u1[index2DtoLinear(sizeImage, i, j + 1)] - u1[index2DtoLinear(sizeImage, i, j - 1)]));
}
else
{
ux[tmpIndex] = 0.0f;
}
uxut[tmpIndex] = ux[tmpIndex] * ut[tmpIndex];
uyut[tmpIndex] = uy[tmpIndex] * ut[tmpIndex];
c1[tmpIndex] = 1.0f + tau * ux[tmpIndex] * ux[tmpIndex];
c2[tmpIndex] = tau * ux[tmpIndex] * uy[tmpIndex];
c3[tmpIndex] = 1.0f + tau * uy[tmpIndex] * uy[tmpIndex];
teiler[tmpIndex] = 1.0f / (c1[tmpIndex] * c3[tmpIndex] - c2[tmpIndex] * c2[tmpIndex]);
if (nrhs > 6)
{
v1[tmpIndex] = (float)inputV[tmpIndex];
v2[tmpIndex] = (float)inputV[nPx + tmpIndex];
}
else
{
v1[tmpIndex] = 0.0f;
v2[tmpIndex] = 0.0f;
}
Kty1[tmpIndex] = 0.0f;
Kty2[tmpIndex] = 0.0f;
if (nrhs > 7)
{
y11[tmpIndex] = (float)inputY[tmpIndex];
y12[tmpIndex] = (float)inputY[nPx + tmpIndex];
y21[tmpIndex] = (float)inputY[2 * nPx + tmpIndex];
y22[tmpIndex] = (float)inputY[3 * nPx + tmpIndex];
}
else
{
y11[tmpIndex] = 0.0f;
y12[tmpIndex] = 0.0f;
y21[tmpIndex] = 0.0f;
y22[tmpIndex] = 0.0f;
}
Kx11[tmpIndex] = 0.0f;
Kx12[tmpIndex] = 0.0f;
Kx21[tmpIndex] = 0.0f;
Kx22[tmpIndex] = 0.0f;
}
}
int iterations = 0;
while (err > tol && iterations <= maxIterations)
{
++iterations;
if (iterations % 50 == 0)
{
p = 0.0f;
d = 0.0f;
}
//primal step
#pragma omp parallel for reduction(+:p) private(i,j)
for (j = 0; j < sizeImage[1]; ++j)
{
for (i = 0; i < sizeImage[0]; ++i)
{
int tmpIndex = index2DtoLinear(sizeImage, i, j);
float Kty1Old = Kty1[tmpIndex];
float Kty2Old = Kty2[tmpIndex];
//transpose equals -div
Kty1[tmpIndex] = -(dxm(y11, sizeImage, i, j) + dym(y12, sizeImage, i, j));
Kty2[tmpIndex] = -(dxm(y21, sizeImage, i, j) + dym(y22, sizeImage, i, j));
float b1 = v1[tmpIndex] - tau*(Kty1[tmpIndex] + uxut[tmpIndex]);
float b2 = v2[tmpIndex] - tau*(Kty2[tmpIndex] + uyut[tmpIndex]);
float v1Old = v1[tmpIndex];
float v2Old = v2[tmpIndex];
v1[tmpIndex] = (b1 * c3[tmpIndex] - c2[tmpIndex] * b2) * teiler[tmpIndex];
v2[tmpIndex] = (b2 * c1[tmpIndex] - c2[tmpIndex] * b1) * teiler[tmpIndex];
if (iterations % 50 == 0)
{
//residuals
p += myAbs((v1Old - v1[tmpIndex]) * dTau - Kty1Old + Kty1[tmpIndex])
+ myAbs((v2Old - v2[tmpIndex]) * dTau - Kty2Old + Kty2[tmpIndex]);
}
}
}
//dual step
#pragma omp parallel for reduction(+:d) private(i,j)
for (j = 0; j < sizeImage[1]; ++j)
{
for (i = 0; i < sizeImage[0]; ++i)
{
int tmpIndex = index2DtoLinear(sizeImage, i, j);
float Kx11Old = Kx11[tmpIndex];
float Kx12Old = Kx12[tmpIndex];
float Kx21Old = Kx21[tmpIndex];
float Kx22Old = Kx22[tmpIndex];
Kx11[tmpIndex] = dxp(v1, sizeImage, i, j);
Kx12[tmpIndex] = dyp(v1, sizeImage, i, j);
Kx21[tmpIndex] = dxp(v2, sizeImage, i, j);
Kx22[tmpIndex] = dyp(v2, sizeImage, i, j);
float y11Old = y11[tmpIndex];
float y12Old = y12[tmpIndex];
float y21Old = y21[tmpIndex];
float y22Old = y22[tmpIndex];
y11[tmpIndex] = ssl * (y11[tmpIndex] + sigma*(2 * Kx11[tmpIndex] - Kx11Old));
y12[tmpIndex] = ssl * (y12[tmpIndex] + sigma*(2 * Kx12[tmpIndex] - Kx12Old));
y21[tmpIndex] = ssl * (y21[tmpIndex] + sigma*(2 * Kx21[tmpIndex] - Kx21Old));
y22[tmpIndex] = ssl * (y22[tmpIndex] + sigma*(2 * Kx22[tmpIndex] - Kx22Old));
if (iterations % 50 == 0)
{
d += myAbs((y11Old - y11[tmpIndex]) * dSigma - Kx11Old + Kx11[tmpIndex]) +
myAbs((y12Old - y12[tmpIndex]) * dSigma - Kx12Old + Kx12[tmpIndex]) +
myAbs((y21Old - y21[tmpIndex]) * dSigma - Kx21Old + Kx21[tmpIndex]) +
myAbs((y22Old - y22[tmpIndex]) * dSigma - Kx22Old + Kx22[tmpIndex]);
}
}
}
if (iterations % 50 == 0)
{
err = (d*d + p*p) / nPx;
}
if (iterations % 1000 == 0)
{
mexPrintf("Iteration %d,Residual %e\n", iterations, err);
mexEvalString("drawnow;");
}
}
//write output
#pragma omp parallel for private(i,j)
for (j = 0; j < sizeImage[1]; ++j)
{
for (i = 0; i < sizeImage[0]; ++i)
{
int tmpIndex = index2DtoLinear(sizeImage, i, j);
YOut[tmpIndex] = (double)y11[tmpIndex];
YOut[tmpIndex + nPx] = (double)y12[tmpIndex];
YOut[tmpIndex + 2 * nPx] = (double)y21[tmpIndex];
YOut[tmpIndex + 3 * nPx] = (double)y22[tmpIndex];
Outv1[tmpIndex] = (double)v1[tmpIndex];
Outv2[tmpIndex] = (double)v2[tmpIndex];
}
}
delete[] v1;
delete[] v2;
delete[] ux;
delete[] uy;
delete[] ut;
delete[] uxut;
delete[] uyut;
delete[] c1;
delete[] c2;
delete[] c3;
delete[] teiler;
delete[] y11;
delete[] y12;
delete[] y21;
delete[] y22;
delete[] Kty1;
delete[] Kty2;
delete[] Kx11;
delete[] Kx12;
delete[] Kx21;
delete[] Kx22;
}
