void cv::gpu::interpolateFrames(const GpuMat& frame0, const GpuMat& frame1, const GpuMat& fu, const GpuMat& fv, const GpuMat& bu, const GpuMat& bv,
float pos, GpuMat& newFrame, GpuMat& buf, Stream& s)
{
CV_Assert(frame0.type() == CV_32FC1);
CV_Assert(frame1.size() == frame0.size() && frame1.type() == frame0.type());
CV_Assert(fu.size() == frame0.size() && fu.type() == frame0.type());
CV_Assert(fv.size() == frame0.size() && fv.type() == frame0.type());
CV_Assert(bu.size() == frame0.size() && bu.type() == frame0.type());
CV_Assert(bv.size() == frame0.size() && bv.type() == frame0.type());
newFrame.create(frame0.size(), frame0.type());
buf.create(6 * frame0.rows, frame0.cols, CV_32FC1);
buf.setTo(Scalar::all(0));
// occlusion masks
GpuMat occ0 = buf.rowRange(0 * frame0.rows, 1 * frame0.rows);
GpuMat occ1 = buf.rowRange(1 * frame0.rows, 2 * frame0.rows);
// interpolated forward flow
GpuMat fui = buf.rowRange(2 * frame0.rows, 3 * frame0.rows);
GpuMat fvi = buf.rowRange(3 * frame0.rows, 4 * frame0.rows);
// interpolated backward flow
GpuMat bui = buf.rowRange(4 * frame0.rows, 5 * frame0.rows);
GpuMat bvi = buf.rowRange(5 * frame0.rows, 6 * frame0.rows);
size_t step = frame0.step;
CV_Assert(frame1.step == step && fu.step == step && fv.step == step && bu.step == step && bv.step == step && newFrame.step == step && buf.step == step);
cudaStream_t stream = StreamAccessor::getStream(s);
NppStStreamHandler h(stream);
NppStInterpolationState state;
state.size = NcvSize32u(frame0.cols, frame0.rows);
state.nStep = static_cast<Ncv32u>(step);
state.pSrcFrame0 = const_cast<Ncv32f*>(frame0.ptr<Ncv32f>());
state.pSrcFrame1 = const_cast<Ncv32f*>(frame1.ptr<Ncv32f>());
state.pFU = const_cast<Ncv32f*>(fu.ptr<Ncv32f>());
state.pFV = const_cast<Ncv32f*>(fv.ptr<Ncv32f>());
state.pBU = const_cast<Ncv32f*>(bu.ptr<Ncv32f>());
state.pBV = const_cast<Ncv32f*>(bv.ptr<Ncv32f>());
state.pos = pos;
state.pNewFrame = newFrame.ptr<Ncv32f>();
state.ppBuffers[0] = occ0.ptr<Ncv32f>();
state.ppBuffers[1] = occ1.ptr<Ncv32f>();
state.ppBuffers[2] = fui.ptr<Ncv32f>();
state.ppBuffers[3] = fvi.ptr<Ncv32f>();
state.ppBuffers[4] = bui.ptr<Ncv32f>();
state.ppBuffers[5] = bvi.ptr<Ncv32f>();
ncvSafeCall( nppiStInterpolateFrames(&state) );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void cv::gpu::ORB_GPU::mergeKeyPoints(GpuMat& keypoints)
{
using namespace cv::gpu::device::orb;
int nAllkeypoints = 0;
for (int level = 0; level < nLevels_; ++level)
nAllkeypoints += keyPointsCount_[level];
if (nAllkeypoints == 0)
{
keypoints.release();
return;
}
ensureSizeIsEnough(ROWS_COUNT, nAllkeypoints, CV_32FC1, keypoints);
int offset = 0;
for (int level = 0; level < nLevels_; ++level)
{
if (keyPointsCount_[level] == 0)
continue;
float sf = getScale(scaleFactor_, firstLevel_, level);
GpuMat keyPointsRange = keypoints.colRange(offset, offset + keyPointsCount_[level]);
float locScale = level != firstLevel_ ? sf : 1.0f;
mergeLocation_gpu(keyPointsPyr_[level].ptr<short2>(0), keyPointsRange.ptr<float>(0), keyPointsRange.ptr<float>(1), keyPointsCount_[level], locScale, 0);
GpuMat range = keyPointsRange.rowRange(2, 4);
keyPointsPyr_[level](Range(1, 3), Range(0, keyPointsCount_[level])).copyTo(range);
keyPointsRange.row(4).setTo(Scalar::all(level));
keyPointsRange.row(5).setTo(Scalar::all(patchSize_ * sf));
offset += keyPointsCount_[level];
}
}
void cv::gpu::ORB_GPU::computeDescriptors(GpuMat& descriptors)
{
using namespace cv::gpu::device::orb;
int nAllkeypoints = 0;
for (int level = 0; level < nLevels_; ++level)
nAllkeypoints += keyPointsCount_[level];
if (nAllkeypoints == 0)
{
descriptors.release();
return;
}
ensureSizeIsEnough(nAllkeypoints, descriptorSize(), CV_8UC1, descriptors);
int offset = 0;
for (int level = 0; level < nLevels_; ++level)
{
if (keyPointsCount_[level] == 0)
continue;
GpuMat descRange = descriptors.rowRange(offset, offset + keyPointsCount_[level]);
if (blurForDescriptor)
{
// preprocess the resized image
ensureSizeIsEnough(imagePyr_[level].size(), imagePyr_[level].type(), buf_);
blurFilter->apply(imagePyr_[level], buf_, Rect(0, 0, imagePyr_[level].cols, imagePyr_[level].rows));
}
computeOrbDescriptor_gpu(blurForDescriptor ? buf_ : imagePyr_[level], keyPointsPyr_[level].ptr<short2>(0), keyPointsPyr_[level].ptr<float>(2),
keyPointsCount_[level], pattern_.ptr<int>(0), pattern_.ptr<int>(1), descRange, descriptorSize(), WTA_K_, 0);
offset += keyPointsCount_[level];
}
}
static void csbp_operator(StereoConstantSpaceBP& rthis, GpuMat& mbuf, GpuMat& temp, GpuMat& out, const GpuMat& left, const GpuMat& right, GpuMat& disp, Stream& stream)
{
CV_DbgAssert(0 < rthis.ndisp && 0 < rthis.iters && 0 < rthis.levels && 0 < rthis.nr_plane
&& left.rows == right.rows && left.cols == right.cols && left.type() == right.type());
CV_Assert(rthis.levels <= 8 && (left.type() == CV_8UC1 || left.type() == CV_8UC3 || left.type() == CV_8UC4));
const Scalar zero = Scalar::all(0);
cudaStream_t cudaStream = StreamAccessor::getStream(stream);
////////////////////////////////////////////////////////////////////////////////////////////
// Init
int rows = left.rows;
int cols = left.cols;
rthis.levels = min(rthis.levels, int(log((double)rthis.ndisp) / log(2.0)));
int levels = rthis.levels;
// compute sizes
AutoBuffer<int> buf(levels * 3);
int* cols_pyr = buf;
int* rows_pyr = cols_pyr + levels;
int* nr_plane_pyr = rows_pyr + levels;
cols_pyr[0] = cols;
rows_pyr[0] = rows;
nr_plane_pyr[0] = rthis.nr_plane;
for (int i = 1; i < levels; i++)
{
cols_pyr[i] = cols_pyr[i-1] / 2;
rows_pyr[i] = rows_pyr[i-1] / 2;
nr_plane_pyr[i] = nr_plane_pyr[i-1] * 2;
}
GpuMat u[2], d[2], l[2], r[2], disp_selected_pyr[2], data_cost, data_cost_selected;
//allocate buffers
int buffers_count = 10; // (up + down + left + right + disp_selected_pyr) * 2
buffers_count += 2; // data_cost has twice more rows than other buffers, what's why +2, not +1;
buffers_count += 1; // data_cost_selected
mbuf.create(rows * rthis.nr_plane * buffers_count, cols, DataType<T>::type);
data_cost = mbuf.rowRange(0, rows * rthis.nr_plane * 2);
data_cost_selected = mbuf.rowRange(data_cost.rows, data_cost.rows + rows * rthis.nr_plane);
for(int k = 0; k < 2; ++k) // in/out
{
GpuMat sub1 = mbuf.rowRange(data_cost.rows + data_cost_selected.rows, mbuf.rows);
GpuMat sub2 = sub1.rowRange((k+0)*sub1.rows/2, (k+1)*sub1.rows/2);
GpuMat *buf_ptrs[] = { &u[k], &d[k], &l[k], &r[k], &disp_selected_pyr[k] };
for(int _r = 0; _r < 5; ++_r)
{
*buf_ptrs[_r] = sub2.rowRange(_r * sub2.rows/5, (_r+1) * sub2.rows/5);
assert(buf_ptrs[_r]->cols == cols && buf_ptrs[_r]->rows == rows * rthis.nr_plane);
}
};
size_t elem_step = mbuf.step / sizeof(T);
Size temp_size = data_cost.size();
if ((size_t)temp_size.area() < elem_step * rows_pyr[levels - 1] * rthis.ndisp)
temp_size = Size(static_cast<int>(elem_step), rows_pyr[levels - 1] * rthis.ndisp);
temp.create(temp_size, DataType<T>::type);
////////////////////////////////////////////////////////////////////////////
// Compute
load_constants(rthis.ndisp, rthis.max_data_term, rthis.data_weight, rthis.max_disc_term, rthis.disc_single_jump, rthis.min_disp_th, left, right, temp);
if (stream)
{
stream.enqueueMemSet(l[0], zero);
stream.enqueueMemSet(d[0], zero);
stream.enqueueMemSet(r[0], zero);
stream.enqueueMemSet(u[0], zero);
stream.enqueueMemSet(l[1], zero);
stream.enqueueMemSet(d[1], zero);
stream.enqueueMemSet(r[1], zero);
stream.enqueueMemSet(u[1], zero);
stream.enqueueMemSet(data_cost, zero);
stream.enqueueMemSet(data_cost_selected, zero);
}
else
{
l[0].setTo(zero);
d[0].setTo(zero);
r[0].setTo(zero);
u[0].setTo(zero);
l[1].setTo(zero);
d[1].setTo(zero);
r[1].setTo(zero);
u[1].setTo(zero);
data_cost.setTo(zero);
data_cost_selected.setTo(zero);
}
int cur_idx = 0;
for (int i = levels - 1; i >= 0; i--)
{
if (i == levels - 1)
{
init_data_cost(left.rows, left.cols, disp_selected_pyr[cur_idx].ptr<T>(), data_cost_selected.ptr<T>(),
elem_step, rows_pyr[i], cols_pyr[i], i, nr_plane_pyr[i], rthis.ndisp, left.channels(), rthis.use_local_init_data_cost, cudaStream);
}
else
{
compute_data_cost(disp_selected_pyr[cur_idx].ptr<T>(), data_cost.ptr<T>(), elem_step,
left.rows, left.cols, rows_pyr[i], cols_pyr[i], rows_pyr[i+1], i, nr_plane_pyr[i+1], left.channels(), cudaStream);
int new_idx = (cur_idx + 1) & 1;
init_message(u[new_idx].ptr<T>(), d[new_idx].ptr<T>(), l[new_idx].ptr<T>(), r[new_idx].ptr<T>(),
u[cur_idx].ptr<T>(), d[cur_idx].ptr<T>(), l[cur_idx].ptr<T>(), r[cur_idx].ptr<T>(),
disp_selected_pyr[new_idx].ptr<T>(), disp_selected_pyr[cur_idx].ptr<T>(),
data_cost_selected.ptr<T>(), data_cost.ptr<T>(), elem_step, rows_pyr[i],
cols_pyr[i], nr_plane_pyr[i], rows_pyr[i+1], cols_pyr[i+1], nr_plane_pyr[i+1], cudaStream);
cur_idx = new_idx;
}
calc_all_iterations(u[cur_idx].ptr<T>(), d[cur_idx].ptr<T>(), l[cur_idx].ptr<T>(), r[cur_idx].ptr<T>(),
data_cost_selected.ptr<T>(), disp_selected_pyr[cur_idx].ptr<T>(), elem_step,
rows_pyr[i], cols_pyr[i], nr_plane_pyr[i], rthis.iters, cudaStream);
}
if (disp.empty())
disp.create(rows, cols, CV_16S);
out = ((disp.type() == CV_16S) ? disp : (out.create(rows, cols, CV_16S), out));
if (stream)
stream.enqueueMemSet(out, zero);
else
out.setTo(zero);
compute_disp(u[cur_idx].ptr<T>(), d[cur_idx].ptr<T>(), l[cur_idx].ptr<T>(), r[cur_idx].ptr<T>(),
data_cost_selected.ptr<T>(), disp_selected_pyr[cur_idx].ptr<T>(), elem_step, out, nr_plane_pyr[0], cudaStream);
if (disp.type() != CV_16S)
{
if (stream)
stream.enqueueConvert(out, disp, disp.type());
else
out.convertTo(disp, disp.type());
}
}