FlowSupportOp(f64 ti, f64 kti, f64 kni, RelaxFlowOpParams ¶ms)
: params_(params),
kernels_(params.numOrientations,
params.numCurvatures,
params.numCurvatures),
kernBufs_(params.numOrientations,
params.numCurvatures,
params.numCurvatures),
ti_(ti),
kti_(kti),
kni_(kni)
{
for (i32 i = 0; i < kernels_.numElems(); i++)
kernels_[i] = ImageData(params.kernelSize, params.kernelSize);
}
virtual void calculateConnections() {
f32 maxExcitation = 0.f;
i32 nto = params_.numOrientations;
i32 nk = params_.numCurvatures;
i32 kernSize = params_.kernelSize;
i32 radius = kernSize/2;
i32 radSquared = radius*radius;
f32 dt = params_.orientationStep;
f32 dk = params_.curvatureStep;
i32 nsamps = params_.subsamples;
f64 sBndry = f64(nsamps/2)/nsamps;
f64 sIncr = 1./nsamps;
f64 stBndry = sBndry*dt;
f64 stIncr = sIncr*dt;
f64 skBndry = sBndry*dk;
f64 skIncr = sIncr*dk;
// The following five for-loops implement subsampling in each dimension
for (f64 sx = -sBndry; sx <= sBndry + sIncr/2; sx += sIncr) {
for (f64 sy = -sBndry; sy <= sBndry + sIncr/2; sy += sIncr) {
for (f64 st = -stBndry; st <= stBndry + stIncr/2; st += stIncr) {
for (f64 skt = -skBndry; skt <= skBndry + skIncr/2; skt += skIncr) {
for (f64 skn = -skBndry; skn <= skBndry + skIncr/2; skn += skIncr) {
model_.set(sx, sy, ti_ + st, kti_ + skt, kni_ + skn);
for (i32 yi = 0; yi < kernSize; yi++) {
i32 yoff = yi - radius;
f64 y = yoff + sy;
for (i32 xi = 0; xi < kernSize; xi++) {
i32 xoff = xi - radius;
if (xoff*xoff + yoff*yoff > radSquared) {
continue;
}
f64 x = xoff + sx;
f64 theta, kt, kn;
if (!model_.valuesAt(x, y, &theta, &kt, &kn)) {
continue;
}
i32 tji = i32(round(theta/dt));
if (params_.numPis == 1) {
if (tji < 0) {
tji += nto;
kt = -kt;
kn = -kn;
}
else if (tji >= nto) {
tji -= nto;
kt = -kt;
kn = -kn;
}
}
tji = clamp(tji, 0, nto - 1);
i32 ktji = i32(round(kt/dk));
ktji = clamp(ktji + nk/2, 0, nk - 1);
i32 knji = i32(round(kn/dk));
knji = clamp(knji + nk/2, 0, nk - 1);
f32& kernVal = kernels_(tji, ktji, knji)(xi, yi);
kernVal += 1;
maxExcitation += 1;
}
}
}}}}}
for (i32 i = 0; i < kernels_.numElems(); ++i)
kernels_[i].data() /= maxExcitation;
}
void initSparseComponents() {
for (i32 i = 0; i < components_.numElems(); i++)
sparseComponents_[i] = SparseImageData(components_[i]);
}
