void cv::gpu::normalize(const GpuMat& src, GpuMat& dst, double a, double b, int norm_type, int dtype, const GpuMat& mask, GpuMat& norm_buf, GpuMat& cvt_buf)
{
double scale = 1, shift = 0;
if (norm_type == NORM_MINMAX)
{
double smin = 0, smax = 0;
double dmin = std::min(a, b), dmax = std::max(a, b);
minMax(src, &smin, &smax, mask, norm_buf);
scale = (dmax - dmin) * (smax - smin > numeric_limits<double>::epsilon() ? 1.0 / (smax - smin) : 0.0);
shift = dmin - smin * scale;
}
else if (norm_type == NORM_L2 || norm_type == NORM_L1 || norm_type == NORM_INF)
{
scale = norm(src, norm_type, mask, norm_buf);
scale = scale > numeric_limits<double>::epsilon() ? a / scale : 0.0;
shift = 0;
}
else
{
CV_Error(CV_StsBadArg, "Unknown/unsupported norm type");
}
if (mask.empty())
{
src.convertTo(dst, dtype, scale, shift);
}
else
{
src.convertTo(cvt_buf, dtype, scale, shift);
cvt_buf.copyTo(dst, mask);
}
}
void cv::gpu::PyrLKOpticalFlow::dense(const GpuMat& prevImg, const GpuMat& nextImg, GpuMat& u, GpuMat& v, GpuMat* err)
{
using namespace cv::gpu::device::pyrlk;
CV_Assert(prevImg.type() == CV_8UC1);
CV_Assert(prevImg.size() == nextImg.size() && prevImg.type() == nextImg.type());
CV_Assert(maxLevel >= 0);
CV_Assert(winSize.width > 2 && winSize.height > 2);
if (err)
err->create(prevImg.size(), CV_32FC1);
// build the image pyramids.
prevPyr_.resize(maxLevel + 1);
nextPyr_.resize(maxLevel + 1);
prevPyr_[0] = prevImg;
nextImg.convertTo(nextPyr_[0], CV_32F);
for (int level = 1; level <= maxLevel; ++level)
{
pyrDown(prevPyr_[level - 1], prevPyr_[level]);
pyrDown(nextPyr_[level - 1], nextPyr_[level]);
}
uPyr_.resize(2);
vPyr_.resize(2);
ensureSizeIsEnough(prevImg.size(), CV_32FC1, uPyr_[0]);
ensureSizeIsEnough(prevImg.size(), CV_32FC1, vPyr_[0]);
ensureSizeIsEnough(prevImg.size(), CV_32FC1, uPyr_[1]);
ensureSizeIsEnough(prevImg.size(), CV_32FC1, vPyr_[1]);
uPyr_[1].setTo(Scalar::all(0));
vPyr_[1].setTo(Scalar::all(0));
int2 winSize2i = make_int2(winSize.width, winSize.height);
loadConstants(winSize2i, iters);
DevMem2Df derr = err ? *err : DevMem2Df();
int idx = 0;
for (int level = maxLevel; level >= 0; level--)
{
int idx2 = (idx + 1) & 1;
lkDense_gpu(prevPyr_[level], nextPyr_[level], uPyr_[idx], vPyr_[idx], uPyr_[idx2], vPyr_[idx2],
level == 0 ? derr : DevMem2Df(), winSize2i);
if (level > 0)
idx = idx2;
}
uPyr_[idx].copyTo(u);
vPyr_[idx].copyTo(v);
}
__host__ GpuMat_<T>::GpuMat_(const GpuMat& m, Allocator* allocator)
: GpuMat(allocator)
{
flags = (flags & ~CV_MAT_TYPE_MASK) | DataType<T>::type;
if (DataType<T>::type == m.type())
{
GpuMat::operator =(m);
return;
}
if (DataType<T>::depth == m.depth())
{
GpuMat::operator =(m.reshape(DataType<T>::channels, m.rows));
return;
}
CV_Assert( DataType<T>::channels == m.channels() );
m.convertTo(*this, type());
}
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());
}
}
static void csbp_operator(StereoConstantSpaceBP& rthis, GpuMat u[2], GpuMat d[2], GpuMat l[2], GpuMat r[2],
GpuMat disp_selected_pyr[2], GpuMat& data_cost, GpuMat& data_cost_selected,
GpuMat& temp, GpuMat& out, const GpuMat& left, const GpuMat& right, GpuMat& disp, cudaStream_t 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));
const Scalar zero = Scalar::all(0);
////////////////////////////////////////////////////////////////////////////////////////////
// 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;
AutoBuffer<int> buf(levels * 4);
int* cols_pyr = buf;
int* rows_pyr = cols_pyr + levels;
int* nr_plane_pyr = rows_pyr + levels;
int* step_pyr = nr_plane_pyr + levels;
cols_pyr[0] = cols;
rows_pyr[0] = rows;
nr_plane_pyr[0] = rthis.nr_plane;
const int n = 64;
step_pyr[0] = alignSize(cols * sizeof(T), n) / sizeof(T);
for (int i = 1; i < levels; i++)
{
cols_pyr[i] = (cols_pyr[i-1] + 1) / 2;
rows_pyr[i] = (rows_pyr[i-1] + 1) / 2;
nr_plane_pyr[i] = nr_plane_pyr[i-1] * 2;
step_pyr[i] = alignSize(cols_pyr[i] * sizeof(T), n) / sizeof(T);
}
Size msg_size(step_pyr[0], rows * nr_plane_pyr[0]);
Size data_cost_size(step_pyr[0], rows * nr_plane_pyr[0] * 2);
u[0].create(msg_size, DataType<T>::type);
d[0].create(msg_size, DataType<T>::type);
l[0].create(msg_size, DataType<T>::type);
r[0].create(msg_size, DataType<T>::type);
u[1].create(msg_size, DataType<T>::type);
d[1].create(msg_size, DataType<T>::type);
l[1].create(msg_size, DataType<T>::type);
r[1].create(msg_size, DataType<T>::type);
disp_selected_pyr[0].create(msg_size, DataType<T>::type);
disp_selected_pyr[1].create(msg_size, DataType<T>::type);
data_cost.create(data_cost_size, DataType<T>::type);
data_cost_selected.create(msg_size, DataType<T>::type);
step_pyr[0] = data_cost.step / sizeof(T);
Size temp_size = data_cost_size;
if (data_cost_size.width * data_cost_size.height < step_pyr[levels - 1] * rows_pyr[levels - 1] * rthis.ndisp)
temp_size = Size(step_pyr[levels - 1], rows_pyr[levels - 1] * rthis.ndisp);
temp.create(temp_size, DataType<T>::type);
////////////////////////////////////////////////////////////////////////////
// Compute
csbp::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);
l[0] = zero;
d[0] = zero;
r[0] = zero;
u[0] = zero;
l[1] = zero;
d[1] = zero;
r[1] = zero;
u[1] = zero;
data_cost = zero;
data_cost_selected = zero;
int cur_idx = 0;
for (int i = levels - 1; i >= 0; i--)
{
if (i == levels - 1)
{
csbp::init_data_cost(left.rows, left.cols, disp_selected_pyr[cur_idx].ptr<T>(), data_cost_selected.ptr<T>(),
step_pyr[i], rows_pyr[i], cols_pyr[i], i, nr_plane_pyr[i], rthis.ndisp, left.channels(), rthis.use_local_init_data_cost, stream);
}
else
{
csbp::compute_data_cost(disp_selected_pyr[cur_idx].ptr<T>(), data_cost.ptr<T>(), step_pyr[i], step_pyr[i+1],
left.rows, left.cols, rows_pyr[i], cols_pyr[i], rows_pyr[i+1], i, nr_plane_pyr[i+1], left.channels(), stream);
int new_idx = (cur_idx + 1) & 1;
csbp::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>(), step_pyr[i], step_pyr[i+1], rows_pyr[i],
cols_pyr[i], nr_plane_pyr[i], rows_pyr[i+1], cols_pyr[i+1], nr_plane_pyr[i+1], stream);
cur_idx = new_idx;
}
csbp::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>(), step_pyr[i],
rows_pyr[i], cols_pyr[i], nr_plane_pyr[i], rthis.iters, stream);
}
if (disp.empty())
disp.create(rows, cols, CV_16S);
out = ((disp.type() == CV_16S) ? disp : (out.create(rows, cols, CV_16S), out));
out = zero;
csbp::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>(), step_pyr[0], out, nr_plane_pyr[0], stream);
if (disp.type() != CV_16S)
out.convertTo(disp, disp.type());
}
void cv::gpu::OpticalFlowDual_TVL1_GPU::operator ()(const GpuMat& I0, const GpuMat& I1, GpuMat& flowx, GpuMat& flowy)
{
CV_Assert( I0.type() == CV_8UC1 || I0.type() == CV_32FC1 );
CV_Assert( I0.size() == I1.size() );
CV_Assert( I0.type() == I1.type() );
CV_Assert( !useInitialFlow || (flowx.size() == I0.size() && flowx.type() == CV_32FC1 && flowy.size() == flowx.size() && flowy.type() == flowx.type()) );
CV_Assert( nscales > 0 );
// allocate memory for the pyramid structure
I0s.resize(nscales);
I1s.resize(nscales);
u1s.resize(nscales);
u2s.resize(nscales);
I0.convertTo(I0s[0], CV_32F, I0.depth() == CV_8U ? 1.0 : 255.0);
I1.convertTo(I1s[0], CV_32F, I1.depth() == CV_8U ? 1.0 : 255.0);
if (!useInitialFlow)
{
flowx.create(I0.size(), CV_32FC1);
flowy.create(I0.size(), CV_32FC1);
}
u1s[0] = flowx;
u2s[0] = flowy;
I1x_buf.create(I0.size(), CV_32FC1);
I1y_buf.create(I0.size(), CV_32FC1);
I1w_buf.create(I0.size(), CV_32FC1);
I1wx_buf.create(I0.size(), CV_32FC1);
I1wy_buf.create(I0.size(), CV_32FC1);
grad_buf.create(I0.size(), CV_32FC1);
rho_c_buf.create(I0.size(), CV_32FC1);
p11_buf.create(I0.size(), CV_32FC1);
p12_buf.create(I0.size(), CV_32FC1);
p21_buf.create(I0.size(), CV_32FC1);
p22_buf.create(I0.size(), CV_32FC1);
diff_buf.create(I0.size(), CV_32FC1);
// create the scales
for (int s = 1; s < nscales; ++s)
{
gpu::resize(I0s[s-1], I0s[s], Size(), scaleStep, scaleStep);
gpu::resize(I1s[s-1], I1s[s], Size(), scaleStep, scaleStep);
if (I0s[s].cols < 16 || I0s[s].rows < 16)
{
nscales = s;
break;
}
if (useInitialFlow)
{
gpu::resize(u1s[s-1], u1s[s], Size(), scaleStep, scaleStep);
gpu::resize(u2s[s-1], u2s[s], Size(), scaleStep, scaleStep);
gpu::multiply(u1s[s], Scalar::all(scaleStep), u1s[s]);
gpu::multiply(u2s[s], Scalar::all(scaleStep), u2s[s]);
}
else
{
u1s[s].create(I0s[s].size(), CV_32FC1);
u2s[s].create(I0s[s].size(), CV_32FC1);
}
}
if (!useInitialFlow)
{
u1s[nscales-1].setTo(Scalar::all(0));
u2s[nscales-1].setTo(Scalar::all(0));
}
// pyramidal structure for computing the optical flow
for (int s = nscales - 1; s >= 0; --s)
{
// compute the optical flow at the current scale
procOneScale(I0s[s], I1s[s], u1s[s], u2s[s]);
// if this was the last scale, finish now
if (s == 0)
break;
// otherwise, upsample the optical flow
// zoom the optical flow for the next finer scale
gpu::resize(u1s[s], u1s[s - 1], I0s[s - 1].size());
gpu::resize(u2s[s], u2s[s - 1], I0s[s - 1].size());
// scale the optical flow with the appropriate zoom factor
gpu::multiply(u1s[s - 1], Scalar::all(1/scaleStep), u1s[s - 1]);
gpu::multiply(u2s[s - 1], Scalar::all(1/scaleStep), u2s[s - 1]);
}
}
void cv::gpu::PyrLKOpticalFlow::sparse(const GpuMat& prevImg, const GpuMat& nextImg, const GpuMat& prevPts, GpuMat& nextPts, GpuMat& status, GpuMat* err)
{
using namespace cv::gpu::device::pyrlk;
if (prevPts.empty())
{
nextPts.release();
status.release();
if (err) err->release();
return;
}
dim3 block, patch;
calcPatchSize(winSize, block, patch, isDeviceArch11_);
CV_Assert(prevImg.type() == CV_8UC1 || prevImg.type() == CV_8UC3 || prevImg.type() == CV_8UC4);
CV_Assert(prevImg.size() == nextImg.size() && prevImg.type() == nextImg.type());
CV_Assert(maxLevel >= 0);
CV_Assert(winSize.width > 2 && winSize.height > 2);
CV_Assert(patch.x > 0 && patch.x < 6 && patch.y > 0 && patch.y < 6);
CV_Assert(prevPts.rows == 1 && prevPts.type() == CV_32FC2);
if (useInitialFlow)
CV_Assert(nextPts.size() == prevPts.size() && nextPts.type() == CV_32FC2);
else
ensureSizeIsEnough(1, prevPts.cols, prevPts.type(), nextPts);
GpuMat temp1 = (useInitialFlow ? nextPts : prevPts).reshape(1);
GpuMat temp2 = nextPts.reshape(1);
multiply(temp1, Scalar::all(1.0 / (1 << maxLevel) / 2.0), temp2);
ensureSizeIsEnough(1, prevPts.cols, CV_8UC1, status);
status.setTo(Scalar::all(1));
if (err)
ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, *err);
// build the image pyramids.
prevPyr_.resize(maxLevel + 1);
nextPyr_.resize(maxLevel + 1);
int cn = prevImg.channels();
if (cn == 1 || cn == 4)
{
prevImg.convertTo(prevPyr_[0], CV_32F);
nextImg.convertTo(nextPyr_[0], CV_32F);
}
else
{
cvtColor(prevImg, dx_calcBuf_, COLOR_BGR2BGRA);
dx_calcBuf_.convertTo(prevPyr_[0], CV_32F);
cvtColor(nextImg, dx_calcBuf_, COLOR_BGR2BGRA);
dx_calcBuf_.convertTo(nextPyr_[0], CV_32F);
}
for (int level = 1; level <= maxLevel; ++level)
{
pyrDown(prevPyr_[level - 1], prevPyr_[level]);
pyrDown(nextPyr_[level - 1], nextPyr_[level]);
}
loadConstants(make_int2(winSize.width, winSize.height), iters);
for (int level = maxLevel; level >= 0; level--)
{
if (cn == 1)
{
lkSparse1_gpu(prevPyr_[level], nextPyr_[level],
prevPts.ptr<float2>(), nextPts.ptr<float2>(), status.ptr(), level == 0 && err ? err->ptr<float>() : 0, prevPts.cols,
level, block, patch);
}
else
{
lkSparse4_gpu(prevPyr_[level], nextPyr_[level],
prevPts.ptr<float2>(), nextPts.ptr<float2>(), status.ptr(), level == 0 && err ? err->ptr<float>() : 0, prevPts.cols,
level, block, patch);
}
}
}
inline
void Stream::enqueueConvert(const GpuMat& src, OutputArray dst, int dtype, double alpha, double beta)
{
src.convertTo(dst, dtype, alpha, beta, *this);
}