static void processFrameGPU(cv::gpu::GpuMat &m)
{
static cv::gpu::GpuMat aux;
static cv::gpu::GpuMat aux2;
static cv::gpu::GpuMat aux3;
cv::gpu::cvtColor(m, aux, CV_BGR2GRAY);
cv::gpu::GaussianBlur(aux,aux2, cv::Size(7,7),1.5);
cv::gpu::Canny(aux2, aux3, 2, 15);
aux3.copyTo(m);
}
void cv::softcascade::SCascade::detect(InputArray _image, InputArray _rois, OutputArray _objects, cv::gpu::Stream& s) const
{
CV_Assert(fields);
// only color images and precomputed integrals are supported
int type = _image.type();
CV_Assert(type == CV_8UC3 || type == CV_32SC1 || (!_rois.empty()));
const cv::gpu::GpuMat image = _image.getGpuMat();
if (_objects.empty()) _objects.create(1, 4096 * sizeof(Detection), CV_8UC1);
cv::gpu::GpuMat rois = _rois.getGpuMat(), objects = _objects.getGpuMat();
/// roi
Fields& flds = *fields;
int shr = flds.shrinkage;
flds.mask.create( rois.cols / shr, rois.rows / shr, rois.type());
device::shrink(rois, flds.mask);
//cv::gpu::transpose(flds.genRoiTmp, flds.mask, s);
if (type == CV_8UC3)
{
flds.update(image.rows, image.cols, flds.shrinkage);
if (flds.check((float)minScale, (float)maxScale, scales))
flds.createLevels(image.rows, image.cols);
flds.preprocessor->apply(image, flds.shrunk);
integral(flds.shrunk, flds.hogluv, flds.integralBuffer, s);
}
else
{
if (s)
s.enqueueCopy(image, flds.hogluv);
else
image.copyTo(flds.hogluv);
}
flds.detect(objects, s);
if ( (flags && NMS_MASK) != NO_REJECT)
{
cv::gpu::GpuMat spr(objects, cv::Rect(0, 0, flds.suppressed.cols, flds.suppressed.rows));
flds.suppress(objects, s);
flds.suppressed.copyTo(spr);
}
}
void LukasKanadeOpticalFlow::drawMotionField_GPU(cv::gpu::GpuMat &imgU, cv::gpu::GpuMat &imgV, cv::gpu::GpuMat &imgMotion,
int xSpace, int ySpace, float minCutoff, float maxCutoff, float multiplier, CvScalar color)
{
cv::Mat uMat( imgU );
cv::Mat vMat( imgV );
cv::Mat drawMat(cv::Size( imgU.size().width, imgU.size().height), CV_8UC3 );
int x = 0, y = 0;
float *ptri;
float deltaX = 0.0, deltaY = 0.0, angle = 0.0, hyp = 0.0;
cv::Point p0, p1;
for( y = ySpace; y < uMat.rows; y += ySpace )
{
for(x = xSpace; x < uMat.cols; x += xSpace )
{
p0.x = x;
p0.y = y;
ptri = uMat.ptr<float>(y);
deltaX = ptri[x];
ptri = vMat.ptr<float>(y);
deltaY = ptri[x];
angle = atan2(deltaY, deltaX);
hyp = sqrt(deltaX*deltaX + deltaY*deltaY);
if( hyp > minCutoff && hyp < maxCutoff )
{
p1.x = p0.x + cvRound(multiplier*hyp*cos(angle));
p1.y = p0.y + cvRound(multiplier*hyp*sin(angle));
cv::line(drawMat,p0,p1,color,1,CV_AA,0);
/*
p0.x = p1.x + cvRound(2*cos(angle-M_PI + M_PI/4));
p0.y = p1.y + cvRound(2*sin(angle-M_PI + M_PI/4));
cv::line( imgMotion, p0, p1, color,1, CV_AA, 0);
p0.x = p1.x + cvRound(2*cos(angle-M_PI - M_PI/4));
p0.y = p1.y + cvRound(2*sin(angle-M_PI - M_PI/4));
cv::line( imgMotion, p0, p1, color,1, CV_AA, 0);
*/
}
}
}
imgMotion.upload( drawMat );
}
void sg::uncropFFTGPU( cv::gpu::GpuMat const & input, cv::gpu::GpuMat & output, std::vector<cv::gpu::GpuMat> & splitBuffer )
{
cv::Size originalSize = input.size();
cv::Size oldSize = output.size();
if( input.channels() > 1 )
{
cv::gpu::split( input, splitBuffer );
splitBuffer[0]( cv::Rect( 0, 0, oldSize.width, oldSize.height ) ).copyTo( output );
}
else
input( cv::Rect( 0, 0, oldSize.width, oldSize.height ) ).copyTo( output );
cv::gpu::multiply( output, ( 1.0 / ( originalSize.width * originalSize.height ) ), output );
}
void MotionSubtraction::subtractGlobalMotion(cv::gpu::GpuMat &in__flowVector3DAngle, cv::gpu::GpuMat &in__flowVector3DMagnitude, cv::gpu::GpuMat &in__globalMotionX, cv::gpu::GpuMat &in__globalMotionY, cv::gpu::GpuMat &out__subtractedAngle, cv::gpu::GpuMat &out__subtractedMagnitude) {
const int64 start = cv::getTickCount();
//----------------------------------------------------------------------------------------
// convert from magnitde/angle to x/y
//----------------------------------------------------------------------------------------
cv::gpu::GpuMat subtractedX(in__flowVector3DAngle.size(), CV_32FC1, cv::Scalar(0.0f));
cv::gpu::GpuMat subtractedY(in__flowVector3DAngle.size(), CV_32FC1, cv::Scalar(0.0f));
cv::gpu::GpuMat flowVector3DX, flowVector3DY;
cv::gpu::polarToCart(in__flowVector3DMagnitude, in__flowVector3DAngle, flowVector3DX, flowVector3DY, true);
//---------------------------------------------------------------------------------------------------------------------------------
// segmentation
//---------------------------------------------------------------------------------------------------------------------------------
kernel.subtractMotion(flowVector3DX, flowVector3DY, in__globalMotionX, in__globalMotionY, subtractedX, subtractedY);
//----------------------------------------------------------------------------------------
// convert from x/y to magnitde/angle
//----------------------------------------------------------------------------------------
cv::gpu::cartToPolar(subtractedX, subtractedY, out__subtractedMagnitude, out__subtractedAngle, true);
//---------------------------------------------------------------------------------------------------------------------------------
// display computation time
//---------------------------------------------------------------------------------------------------------------------------------
const double timeSec = (cv::getTickCount() - start) / cv::getTickFrequency();
std::cout << "Motion subtr : \t" << timeSec << " sec" << std::endl;
}
void buildPyramid(const cv::gpu::GpuMat& cvgmGray_){
cvgmGray_.copyTo(*_acvgmShrPtrPyrBWs[0]);
for (int n=0; n< 3; n++){
cv::gpu::resize(*_acvgmShrPtrPyrBWs[n],*_acvgmShrPtrPyrBWs[n+1],cv::Size(0,0),.5f,.5f );
}
return;
}
void CFast::operator ()(const cv::gpu::GpuMat& cvgmImage_, const cv::gpu::GpuMat& cvgmMask_, std::vector<cv::KeyPoint>* pvKeyPoints_)
{
if (cvgmImage_.empty())
return;
(*this)(cvgmImage_, cvgmMask_, &_cvgmdKeyPoints);
downloadKeypoints(_cvgmdKeyPoints, &*pvKeyPoints_);
}
//! find keypoints and compute it's response if _bNonMaxSupression is true
//! return count of detected keypoints
//return the # of keypoints, stored in _uCount;
//store Key point locations into _cvgmKeyPointLocation;
//store the corner strength into _cvgmScore;
int CFast::calcKeyPointsLocation(const cv::gpu::GpuMat& cvgmImg_, const cv::gpu::GpuMat& cvgmMask_)
{
//using namespace cv::gpu::device::fast;
CV_Assert(cvgmImg_.type() == CV_8UC1);
CV_Assert(cvgmMask_.empty() || (cvgmMask_.type() == CV_8UC1 && cvgmMask_.size() == cvgmImg_.size()));
if (!cv::gpu::TargetArchs::builtWith(cv::gpu::GLOBAL_ATOMICS) || !cv::gpu::DeviceInfo().supports(cv::gpu::GLOBAL_ATOMICS))
CV_Error(CV_StsNotImplemented, "The device doesn't support global atomics");
unsigned int uMaxKeypoints = static_cast<unsigned int>(_dKeyPointsRatio * cvgmImg_.size().area());
ensureSizeIsEnough(1, uMaxKeypoints, CV_16SC2, _cvgmKeyPointLocation);
if (_bNonMaxSupression)
{
ensureSizeIsEnough(cvgmImg_.size(), CV_32SC1, _cvgmScore);
_cvgmScore.setTo(cv::Scalar::all(0));
}
_uCount = btl::device::fast::cudaCalcKeypoints(cvgmImg_, cvgmMask_, uMaxKeypoints, _nThreshold, _cvgmKeyPointLocation.ptr<short2>(), _bNonMaxSupression ? &_cvgmScore : NULL);
_uCount = std::min(_uCount, uMaxKeypoints);
return _uCount;
}
void Tester::testDepth(cv::gpu::GpuMat &testDepth, cv::gpu::GpuMat &groundTruthMap, float &differenceValue) {
// download depthmaps
cv::Mat depth, groundTruth;
testDepth.download(depth);
groundTruthMap.download(groundTruth);
float sum = 0.0f;
for (int y = 0; y < depth.rows; y++) {
for (int x = 0; x < depth.cols; x++) {
sum = (groundTruth.at<float>(y,x) - depth.at<float>(y,x)) * (groundTruth.at<float>(y,x) - depth.at<float>(y,x));
}
}
differenceValue = sum;
file << "Depth error: " << differenceValue << "\n";
}
void meanStdDev(const cv::gpu::GpuMat& mtx,cv::Scalar& mean, cv::Scalar& stddev){
assert(1==mtx.channels());
size_t elem_num = mtx.rows * mtx.cols;
cv::gpu::GpuMat buf;
mean = cv::gpu::sum( mtx, buf);
stddev = cv::gpu::sqrSum( mtx, buf);
for(int i=0;i<mean.cols;i++){
mean[i] /= elem_num;
}
stddev = cv::gpu::sqrSum( mtx, buf);
for(int i=0;i<stddev.cols;i++){
stddev[i] /= elem_num;
stddev[i] = std::sqrt(stddev[i] - std::pow(mean[i],2));
}
}
void sg::DoGCenterSurround::operator()( cv::gpu::GpuMat const & input, cv::gpu::GpuMat & output )
{
if( !itsIsValid )
return;
// Zero the output
output.setTo( 0.0f );
for( auto & filter : itsFilters )
{
cv::gpu::mulSpectrums( input, filter, itsBuffer, cv::DFT_COMPLEX_OUTPUT );
addFC2Wrapper( output, itsBuffer, output );
}
mulValueFC2Wrapper( output, 1.0f / itsFilters.size(), output );
}
void detectKeypoints(cv::gpu::GpuMat& keypoints, int scales)
{
ensureSizeIsEnough(SIFT_GPU::ROWS_COUNT, MAXEXTREMAS, CV_32FC1, keypoints);
keypoints.setTo(cv::Scalar::all(0));
for (int octave = 0; octave < 1; ++octave)
{
const int scaleCols = cols >> octave;
const int scaleRows = rows >> octave;
createDoGSpace(inImage.data, &deviceDoGData, scales, scaleRows, scaleCols);
findExtremas(deviceDoGData, &sift_.extremaBuffer, &maxCounter, octave, scales, scaleRows, scaleCols);
localization(deviceDoGData, scaleRows, scaleCols, scales, octave, sift_.nOctaves, sift_.extremaBuffer, maxCounter,
keypoints.ptr<float>(SIFT_GPU::X_ROW), keypoints.ptr<float>(SIFT_GPU::Y_ROW),
keypoints.ptr<float>(SIFT_GPU::OCTAVE_ROW), keypoints.ptr<float>(SIFT_GPU::SIZE_ROW),
keypoints.ptr<float>(SIFT_GPU::ANGLE_ROW), keypoints.ptr<float>(SIFT_GPU::RESPONSE_ROW));
}
std::cout << "Number of keypoints: " << maxCounter[0] << std::endl;
}
void sg::LogGabor::getEdgeResponses( cv::gpu::GpuMat const & fftImage, cv::gpu::GpuMat & edges,
std::vector<cv::gpu::GpuMat> & splitBuffer, DoGCenterSurround & dog )
{
if( !itsValid )
{
std::cerr << "This LogGabor bank is not initialized!\n";
return;
}
edges.setTo( 0.0f );
// outer loop for orientations
for( size_t o = 0; o < itsFilters.size(); ++o )
{
// inner loop over scales
for( size_t s = 0; s < itsGaborScales.size(); ++s )
{
// see note in addFilters about why this is a normal multiplication
// and not a spectrum multiply
mulFC2Wrapper( fftImage, itsFilters[o][s], itsFFTBuffer[0] );
cv::gpu::dft( itsFFTBuffer[0], itsFFTBuffer[1], itsFFTBuffer[0].size(), cv::DFT_INVERSE );
// Get magnitude
cv::gpu::magnitude( itsFFTBuffer[1], splitBuffer[0] );
// Compute edge response
// the edge response is a DoG applied to the magnitude
// so we'll pad the magnitude with zero valued complex
// take the DFT and then pass it through the DoG processing chain
splitBuffer[1].setTo( 0.0f );
cv::gpu::GpuMat merge[] = { splitBuffer[0], splitBuffer[1] };
cv::gpu::merge( merge, 2, itsFFTBuffer[0] );
cv::gpu::dft( itsFFTBuffer[0], itsFFTBuffer[0], itsFFTBuffer[0].size() );
dog( itsFFTBuffer[0], itsFFTBuffer[2] );
// accumulate result into edges
addFC2Wrapper( edges, itsFFTBuffer[2], edges );
} // end scales
} // end orientations
}
void cv::gpu::MOG_GPU::operator()(const cv::gpu::GpuMat& frame, cv::gpu::GpuMat& fgmask, float learningRate, Stream& stream)
{
using namespace cv::gpu::cudev::mog;
CV_Assert(frame.depth() == CV_8U);
int ch = frame.channels();
int work_ch = ch;
if (nframes_ == 0 || learningRate >= 1.0 || frame.size() != frameSize_ || work_ch != mean_.channels())
initialize(frame.size(), frame.type());
fgmask.create(frameSize_, CV_8UC1);
++nframes_;
learningRate = learningRate >= 0.0f && nframes_ > 1 ? learningRate : 1.0f / std::min(nframes_, history);
CV_Assert(learningRate >= 0.0f);
mog_gpu(frame, ch, fgmask, weight_, sortKey_, mean_, var_, nmixtures_,
varThreshold, learningRate, backgroundRatio, noiseSigma,
StreamAccessor::getStream(stream));
}
void CFast::downloadKeypoints(const cv::gpu::GpuMat& cvgmKeyPoints_, std::vector<cv::KeyPoint>* pvKeyPoints_)
{
if (cvgmKeyPoints_.empty()) return;
cv::Mat cvmKeyPoints(cvgmKeyPoints_);//download from cvgm to cvm
convertKeypoints(cvmKeyPoints, &*pvKeyPoints_);
}
void cv::gpu::GMG_GPU::operator ()(const cv::gpu::GpuMat& frame, cv::gpu::GpuMat& fgmask, float newLearningRate, cv::gpu::Stream& stream)
{
using namespace cv::gpu::cudev::bgfg_gmg;
typedef void (*func_t)(PtrStepSzb frame, PtrStepb fgmask, PtrStepSzi colors, PtrStepf weights, PtrStepi nfeatures,
int frameNum, float learningRate, bool updateBackgroundModel, cudaStream_t stream);
static const func_t funcs[6][4] =
{
{update_gpu<uchar>, 0, update_gpu<uchar3>, update_gpu<uchar4>},
{0,0,0,0},
{update_gpu<ushort>, 0, update_gpu<ushort3>, update_gpu<ushort4>},
{0,0,0,0},
{0,0,0,0},
{update_gpu<float>, 0, update_gpu<float3>, update_gpu<float4>}
};
CV_Assert(frame.depth() == CV_8U || frame.depth() == CV_16U || frame.depth() == CV_32F);
CV_Assert(frame.channels() == 1 || frame.channels() == 3 || frame.channels() == 4);
if (newLearningRate != -1.0f)
{
CV_Assert(newLearningRate >= 0.0f && newLearningRate <= 1.0f);
learningRate = newLearningRate;
}
if (frame.size() != frameSize_)
initialize(frame.size(), 0.0f, frame.depth() == CV_8U ? 255.0f : frame.depth() == CV_16U ? std::numeric_limits<ushort>::max() : 1.0f);
fgmask.create(frameSize_, CV_8UC1);
fgmask.setTo(cv::Scalar::all(0), stream);
funcs[frame.depth()][frame.channels() - 1](frame, fgmask, colors_, weights_, nfeatures_, frameNum_, learningRate, updateBackgroundModel, cv::gpu::StreamAccessor::getStream(stream));
// medianBlur
if (smoothingRadius > 0)
{
boxFilter_->apply(fgmask, buf_, stream);
int minCount = (smoothingRadius * smoothingRadius + 1) / 2;
double thresh = 255.0 * minCount / (smoothingRadius * smoothingRadius);
cv::gpu::threshold(buf_, fgmask, thresh, 255.0, cv::THRESH_BINARY, stream);
}
// keep track of how many frames we have processed
++frameNum_;
}