Skip to content

O
opencv
Commit f196e9fd authored by marina.kolpakova's avatar marina.kolpakova
Browse files
Options

add factory method for Fields structure

parent 0898c3c6
Show whitespace changes
Inline Side-by-side
Showing with 258 additions and 273 deletions
modules/gpu/src/softcascade.cpp
View file @ f196e9fd
......@@ -78,203 +78,63 @@ namespace imgproc
struct cv::gpu::SoftCascade::Filds
{
Filds()
{
plane.create(FRAME_HEIGHT * (HOG_LUV_BINS + 1), FRAME_WIDTH, CV_8UC1);
fplane.create(FRAME_HEIGHT * 6, FRAME_WIDTH, CV_32FC1);
luv.create(FRAME_HEIGHT, FRAME_WIDTH, CV_8UC3);
shrunk.create(FRAME_HEIGHT / 4 * HOG_LUV_BINS, FRAME_WIDTH / 4, CV_8UC1);
integralBuffer.create(1 , (shrunk.rows + 1) * HOG_LUV_BINS * (shrunk.cols + 1), CV_32SC1);
hogluv.create((FRAME_HEIGHT / 4 + 1) * HOG_LUV_BINS, FRAME_WIDTH / 4 + 64, CV_32SC1);
detCounter.create(1,1, CV_32SC1);
}
// scales range
float minScale;
float maxScale;
int origObjWidth;
int origObjHeight;
int downscales;
GpuMat octaves;
GpuMat stages;
GpuMat nodes;
GpuMat leaves;
GpuMat levels;
GpuMat detCounter;
// preallocated buffer 640x480x10 for hogluv + 640x480 got gray
GpuMat plane;
// preallocated buffer for floating point operations
GpuMat fplane;
// temporial mat for cvtColor
GpuMat luv;
// 160x120x10
GpuMat shrunk;
// temporial mat for integrall
GpuMat integralBuffer;
// 161x121x10
GpuMat hogluv;
std::vector<float> scales;
device::icf::CascadeInvoker<device::icf::CascadePolicy> invoker;
static const int shrinkage = 4;
enum { BOOST = 0 };
enum
{
FRAME_WIDTH = 640,
FRAME_HEIGHT = 480,
TOTAL_SCALES = 55,
ORIG_OBJECT_WIDTH = 64,
ORIG_OBJECT_HEIGHT = 128,
HOG_BINS = 6,
LUV_BINS = 3,
HOG_LUV_BINS = 10
};
bool fill(const FileNode &root, const float mins, const float maxs);
void detect(int scale, const cv::gpu::GpuMat& roi, cv::gpu::GpuMat& objects, cudaStream_t stream) const
{
cudaMemset(detCounter.data, 0, detCounter.step * detCounter.rows * sizeof(int));
// device::icf::CascadeInvoker<device::icf::CascadePolicy> invoker(levels, octaves, stages, nodes, leaves);
invoker(roi, hogluv, objects, detCounter, downscales, scale);
}
void preprocess(const cv::gpu::GpuMat& colored)
{
cudaMemset(plane.data, 0, plane.step * plane.rows);
static const int fw = Filds::FRAME_WIDTH;
static const int fh = Filds::FRAME_HEIGHT;
GpuMat gray(plane, cv::Rect(0, fh * Filds::HOG_LUV_BINS, fw, fh));
cv::gpu::cvtColor(colored, gray, CV_BGR2GRAY);
createHogBins(gray);
createLuvBins(colored);
integrate();
}
private:
void calcLevels(const std::vector<device::icf::Octave>& octs,
int frameW, int frameH, int nscales);
typedef std::vector<device::icf::Octave>::const_iterator octIt_t;
int fitOctave(const std::vector<device::icf::Octave>& octs, const float& logFactor) const
{
float minAbsLog = FLT_MAX;
int res = 0;
for (int oct = 0; oct < (int)octs.size(); ++oct)
{
const device::icf::Octave& octave =octs[oct];
float logOctave = ::log(octave.scale);
float logAbsScale = ::fabs(logFactor - logOctave);
if(logAbsScale < minAbsLog)
struct CascadeIntrinsics
{
res = oct;
minAbsLog = logAbsScale;
}
}
return res;
}
static const float lambda = 1.099f, a = 0.89f;
void createHogBins(const cv::gpu::GpuMat& gray)
static float getFor(int channel, float scaling)
{
static const int fw = Filds::FRAME_WIDTH;
static const int fh = Filds::FRAME_HEIGHT;
GpuMat dfdx(fplane, cv::Rect(0, 0, fw, fh));
GpuMat dfdy(fplane, cv::Rect(0, fh, fw, fh));
cv::gpu::Sobel(gray, dfdx, CV_32F, 1, 0, 3, 0.125f);
cv::gpu::Sobel(gray, dfdy, CV_32F, 0, 1, 3, 0.125f);
GpuMat mag(fplane, cv::Rect(0, 2 * fh, fw, fh));
GpuMat ang(fplane, cv::Rect(0, 3 * fh, fw, fh));
cv::gpu::cartToPolar(dfdx, dfdy, mag, ang, true);
// normolize magnitude to uchar interval and angles to 6 bins
GpuMat nmag(fplane, cv::Rect(0, 4 * fh, fw, fh));
GpuMat nang(fplane, cv::Rect(0, 5 * fh, fw, fh));
cv::gpu::multiply(mag, cv::Scalar::all(1.f / ::log(2)), nmag);
cv::gpu::multiply(ang, cv::Scalar::all(1.f / 60.f), nang);
//create uchar magnitude
GpuMat cmag(plane, cv::Rect(0, fh * Filds::HOG_BINS, fw, fh));
nmag.convertTo(cmag, CV_8UC1);
CV_Assert(channel < 10);
device::icf::fillBins(plane, nang, fw, fh, Filds::HOG_BINS);
}
if (fabs(scaling - 1.f) < FLT_EPSILON)
return 1.f;
void createLuvBins(const cv::gpu::GpuMat& colored)
{
static const int fw = Filds::FRAME_WIDTH;
static const int fh = Filds::FRAME_HEIGHT;
// according to R. Benenson, M. Mathias, R. Timofte and L. Van Gool's and Dallal's papers
static const float A[2][2] =
{ //channel <= 6, otherwise
{ 0.89f, 1.f}, // down
{ 1.00f, 1.f} // up
};
cv::gpu::cvtColor(colored, luv, CV_BGR2Luv);
static const float B[2][2] =
{ //channel <= 6, otherwise
{ 1.099f / ::log(2), 2.f}, // down
{ 0.f, 2.f} // up
};
std::vector<GpuMat> splited;
for(int i = 0; i < Filds::LUV_BINS; ++i)
{
splited.push_back(GpuMat(plane, cv::Rect(0, fh * (7 + i), fw, fh)));
}
float a = A[(int)(scaling >= 1)][(int)(channel > 6)];
float b = B[(int)(scaling >= 1)][(int)(channel > 6)];
cv::gpu::split(luv, splited);
// printf("!!! scaling: %f %f %f -> %f\n", scaling, a, b, a * pow(scaling, b));
return a * ::pow(scaling, b);
}
};
void integrate()
{
int fw = Filds::FRAME_WIDTH;
int fh = Filds::FRAME_HEIGHT;
GpuMat channels(plane, cv::Rect(0, 0, fw, fh * Filds::HOG_LUV_BINS));
cv::gpu::resize(channels, shrunk, cv::Size(), 0.25, 0.25, CV_INTER_AREA);
fw /= shrinkage;
fh /= shrinkage;
for(int i = 0; i < Filds::HOG_LUV_BINS; ++i)
static Filds* parseCascade(const FileNode &root, const float mins, const float maxs)
{
GpuMat channel(shrunk, cv::Rect(0, fh * i, fw, fh ));
GpuMat sum(hogluv, cv::Rect(0, (fh + 1) * i, fw + 1, fh + 1));
cv::gpu::integralBuffered(channel, sum, integralBuffer);
}
}
};
bool cv::gpu::SoftCascade::Filds::fill(const FileNode &root, const float mins, const float maxs)
{
using namespace device::icf;
minScale = mins;
maxScale = maxs;
// cascade properties
static const char *const SC_STAGE_TYPE = "stageType";
static const char *const SC_BOOST = "BOOST";
static const char *const SC_FEATURE_TYPE = "featureType";
static const char *const SC_ICF = "ICF";
// only Ada Boost supported
std::string stageTypeStr = (string)root[SC_STAGE_TYPE];
CV_Assert(stageTypeStr == SC_BOOST);
// only HOG-like integral channel features cupported
string featureTypeStr = (string)root[SC_FEATURE_TYPE];
CV_Assert(featureTypeStr == SC_ICF);
static const char *const SC_ORIG_W = "width";
static const char *const SC_ORIG_H = "height";
int origWidth = (int)root[SC_ORIG_W];
CV_Assert(origWidth == ORIG_OBJECT_WIDTH);
int origHeight = (int)root[SC_ORIG_H];
CV_Assert(origHeight == ORIG_OBJECT_HEIGHT);
static const char *const SC_OCTAVES = "octaves";
static const char *const SC_STAGES = "stages";
static const char *const SC_FEATURES = "features";
......@@ -292,28 +152,16 @@ bool cv::gpu::SoftCascade::Filds::fill(const FileNode &root, const float mins, c
static const char * const SC_F_CHANNEL = "channel";
static const char * const SC_F_RECT = "rect";
// only Ada Boost supported
std::string stageTypeStr = (string)root[SC_STAGE_TYPE];
CV_Assert(stageTypeStr == SC_BOOST);
// only HOG-like integral channel features cupported
string featureTypeStr = (string)root[SC_FEATURE_TYPE];
CV_Assert(featureTypeStr == SC_ICF);
origObjWidth = (int)root[SC_ORIG_W];
CV_Assert(origObjWidth == ORIG_OBJECT_WIDTH);
origObjHeight = (int)root[SC_ORIG_H];
CV_Assert(origObjHeight == ORIG_OBJECT_HEIGHT);
FileNode fn = root[SC_OCTAVES];
if (fn.empty()) return false;
using namespace device::icf;
std::vector<Octave> voctaves;
std::vector<float> vstages;
std::vector<Node> vnodes;
std::vector<float> vleaves;
scales.clear();
FileNodeIterator it = fn.begin(), it_end = fn.end();
int feature_offset = 0;
......@@ -327,7 +175,6 @@ bool cv::gpu::SoftCascade::Filds::fill(const FileNode &root, const float mins, c
bool isUPOctave = scale >= 1;
scales.push_back(scale);
ushort nstages = saturate_cast<ushort>((int)fns[SC_OCT_STAGES]);
ushort2 size;
size.x = cvRound(ORIG_OBJECT_WIDTH * scale);
......@@ -395,82 +242,32 @@ bool cv::gpu::SoftCascade::Filds::fill(const FileNode &root, const float mins, c
++octIndex;
}
// upload in gpu memory
octaves.upload(cv::Mat(1, voctaves.size() * sizeof(Octave), CV_8UC1, (uchar*)&(voctaves[0]) ));
CV_Assert(!octaves.empty());
stages.upload(cv::Mat(vstages).reshape(1,1));
CV_Assert(!stages.empty());
nodes.upload(cv::Mat(1, vnodes.size() * sizeof(Node), CV_8UC1, (uchar*)&(vnodes[0]) ));
CV_Assert(!nodes.empty());
leaves.upload(cv::Mat(vleaves).reshape(1,1));
CV_Assert(!leaves.empty());
// compute levels
calcLevels(voctaves, FRAME_WIDTH, FRAME_HEIGHT, TOTAL_SCALES);
CV_Assert(!levels.empty());
invoker = device::icf::CascadeInvoker<device::icf::CascadePolicy>(levels, octaves, stages, nodes, leaves);
return true;
}
cv::Mat hoctaves(1, voctaves.size() * sizeof(Octave), CV_8UC1, (uchar*)&(voctaves[0]));
CV_Assert(!hoctaves.empty());
namespace {
struct CascadeIntrinsics
{
static const float lambda = 1.099f, a = 0.89f;
cv::Mat hstages(cv::Mat(vstages).reshape(1,1));
CV_Assert(!hstages.empty());
static float getFor(int channel, float scaling)
{
CV_Assert(channel < 10);
cv::Mat hnodes(1, vnodes.size() * sizeof(Node), CV_8UC1, (uchar*)&(vnodes[0]) );
CV_Assert(!hnodes.empty());
if (fabs(scaling - 1.f) < FLT_EPSILON)
return 1.f;
// according to R. Benenson, M. Mathias, R. Timofte and L. Van Gool's and Dallal's papers
static const float A[2][2] =
{ //channel <= 6, otherwise
{ 0.89f, 1.f}, // down
{ 1.00f, 1.f} // up
};
static const float B[2][2] =
{ //channel <= 6, otherwise
{ 1.099f / log(2), 2.f}, // down
{ 0.f, 2.f} // up
};
float a = A[(int)(scaling >= 1)][(int)(channel > 6)];
float b = B[(int)(scaling >= 1)][(int)(channel > 6)];
// printf("!!! scaling: %f %f %f -> %f\n", scaling, a, b, a * pow(scaling, b));
return a * pow(scaling, b);
}
};
}
inline void cv::gpu::SoftCascade::Filds::calcLevels(const std::vector<device::icf::Octave>& octs,
int frameW, int frameH, int nscales)
{
CV_Assert(nscales > 1);
using device::icf::Level;
cv::Mat hleaves(cv::Mat(vleaves).reshape(1,1));
CV_Assert(!hleaves.empty());
std::vector<Level> vlevels;
float logFactor = (::log(maxScale) - ::log(minScale)) / (nscales -1);
float logFactor = (::log(maxs) - ::log(mins)) / (TOTAL_SCALES -1);
float scale = minScale;
downscales = 0;
for (int sc = 0; sc < nscales; ++sc)
float scale = mins;
int downscales = 0;
for (int sc = 0; sc < TOTAL_SCALES; ++sc)
{
int width = ::std::max(0.0f, frameW - (origObjWidth * scale));
int height = ::std::max(0.0f, frameH - (origObjHeight * scale));
int width = ::std::max(0.0f, FRAME_WIDTH - (origWidth * scale));
int height = ::std::max(0.0f, FRAME_HEIGHT - (origHeight * scale));
float logScale = ::log(scale);
int fit = fitOctave(octs, logScale);
int fit = fitOctave(voctaves, logScale);
Level level(fit, octs[fit], scale, width, height);
Level level(fit, voctaves[fit], scale, width, height);
level.scaling[0] = CascadeIntrinsics::getFor(0, level.relScale);
level.scaling[1] = CascadeIntrinsics::getFor(9, level.relScale);
......@@ -479,11 +276,11 @@ inline void cv::gpu::SoftCascade::Filds::calcLevels(const std::vector<device::ic
else
{
vlevels.push_back(level);
if (octs[fit].scale < 1) ++downscales;
if (voctaves[fit].scale < 1) ++downscales;
}
if (::fabs(scale - maxScale) < FLT_EPSILON) break;
scale = ::std::min(maxScale, ::expf(::log(scale) + logFactor));
if (::fabs(scale - maxs) < FLT_EPSILON) break;
scale = ::std::min(maxs, ::expf(::log(scale) + logFactor));
// std::cout << "level " << sc
// << " octeve "
......@@ -496,8 +293,201 @@ inline void cv::gpu::SoftCascade::Filds::calcLevels(const std::vector<device::ic
// << (int)vlevels[sc].workRect.x << " " << (int)vlevels[sc].workRect.y << "]" << std::endl;
}
levels.upload(cv::Mat(1, vlevels.size() * sizeof(Level), CV_8UC1, (uchar*)&(vlevels[0]) ));
}
cv::Mat hlevels(1, vlevels.size() * sizeof(Level), CV_8UC1, (uchar*)&(vlevels[0]) );
CV_Assert(!hlevels.empty());
Filds* filds = new Filds(mins, maxs, origWidth, origHeight, shrinkage, downscales,
hoctaves, hstages, hnodes, hleaves, hlevels);
return filds;
}
Filds( const float mins, const float maxs, const int ow, const int oh, const int shr, const int ds,
cv::Mat hoctaves, cv::Mat hstages, cv::Mat hnodes, cv::Mat hleaves, cv::Mat hlevels)
: minScale(mins), maxScale(maxs), origObjWidth(ow), origObjHeight(oh), shrinkage(shr), downscales(ds)
{
plane.create(FRAME_HEIGHT * (HOG_LUV_BINS + 1), FRAME_WIDTH, CV_8UC1);
fplane.create(FRAME_HEIGHT * 6, FRAME_WIDTH, CV_32FC1);
luv.create(FRAME_HEIGHT, FRAME_WIDTH, CV_8UC3);
shrunk.create(FRAME_HEIGHT / shr * HOG_LUV_BINS, FRAME_WIDTH / shr, CV_8UC1);
integralBuffer.create(1 , (shrunk.rows + 1) * HOG_LUV_BINS * (shrunk.cols + 1), CV_32SC1);
hogluv.create((FRAME_HEIGHT / shr + 1) * HOG_LUV_BINS, FRAME_WIDTH / shr + 64, CV_32SC1);
detCounter.create(1,1, CV_32SC1);
octaves.upload(hoctaves);
stages.upload(hstages);
nodes.upload(hnodes);
leaves.upload(hleaves);
levels.upload(hlevels);
invoker = device::icf::CascadeInvoker<device::icf::CascadePolicy>(levels, octaves, stages, nodes, leaves);
}
void detect(int scale, const cv::gpu::GpuMat& roi, cv::gpu::GpuMat& objects, cudaStream_t stream) const
{
cudaMemset(detCounter.data, 0, detCounter.step * detCounter.rows * sizeof(int));
invoker(roi, hogluv, objects, detCounter, downscales, scale);
}
void preprocess(const cv::gpu::GpuMat& colored)
{
cudaMemset(plane.data, 0, plane.step * plane.rows);
static const int fw = Filds::FRAME_WIDTH;
static const int fh = Filds::FRAME_HEIGHT;
GpuMat gray(plane, cv::Rect(0, fh * Filds::HOG_LUV_BINS, fw, fh));
cv::gpu::cvtColor(colored, gray, CV_BGR2GRAY);
createHogBins(gray);
createLuvBins(colored);
integrate();
}
private:
typedef std::vector<device::icf::Octave>::const_iterator octIt_t;
static int fitOctave(const std::vector<device::icf::Octave>& octs, const float& logFactor)
{
float minAbsLog = FLT_MAX;
int res = 0;
for (int oct = 0; oct < (int)octs.size(); ++oct)
{
const device::icf::Octave& octave =octs[oct];
float logOctave = ::log(octave.scale);
float logAbsScale = ::fabs(logFactor - logOctave);
if(logAbsScale < minAbsLog)
{
res = oct;
minAbsLog = logAbsScale;
}
}
return res;
}
void createHogBins(const cv::gpu::GpuMat& gray)
{
static const int fw = Filds::FRAME_WIDTH;
static const int fh = Filds::FRAME_HEIGHT;
GpuMat dfdx(fplane, cv::Rect(0, 0, fw, fh));
GpuMat dfdy(fplane, cv::Rect(0, fh, fw, fh));
cv::gpu::Sobel(gray, dfdx, CV_32F, 1, 0, 3, 0.125f);
cv::gpu::Sobel(gray, dfdy, CV_32F, 0, 1, 3, 0.125f);
GpuMat mag(fplane, cv::Rect(0, 2 * fh, fw, fh));
GpuMat ang(fplane, cv::Rect(0, 3 * fh, fw, fh));
cv::gpu::cartToPolar(dfdx, dfdy, mag, ang, true);
// normolize magnitude to uchar interval and angles to 6 bins
GpuMat nmag(fplane, cv::Rect(0, 4 * fh, fw, fh));
GpuMat nang(fplane, cv::Rect(0, 5 * fh, fw, fh));
cv::gpu::multiply(mag, cv::Scalar::all(1.f / ::log(2)), nmag);
cv::gpu::multiply(ang, cv::Scalar::all(1.f / 60.f), nang);
//create uchar magnitude
GpuMat cmag(plane, cv::Rect(0, fh * Filds::HOG_BINS, fw, fh));
nmag.convertTo(cmag, CV_8UC1);
device::icf::fillBins(plane, nang, fw, fh, Filds::HOG_BINS);
}
void createLuvBins(const cv::gpu::GpuMat& colored)
{
static const int fw = Filds::FRAME_WIDTH;
static const int fh = Filds::FRAME_HEIGHT;
cv::gpu::cvtColor(colored, luv, CV_BGR2Luv);
std::vector<GpuMat> splited;
for(int i = 0; i < Filds::LUV_BINS; ++i)
{
splited.push_back(GpuMat(plane, cv::Rect(0, fh * (7 + i), fw, fh)));
}
cv::gpu::split(luv, splited);
}
void integrate()
{
int fw = Filds::FRAME_WIDTH;
int fh = Filds::FRAME_HEIGHT;
GpuMat channels(plane, cv::Rect(0, 0, fw, fh * Filds::HOG_LUV_BINS));
cv::gpu::resize(channels, shrunk, cv::Size(), 0.25, 0.25, CV_INTER_AREA);
fw /= shrinkage;
fh /= shrinkage;
for(int i = 0; i < Filds::HOG_LUV_BINS; ++i)
{
GpuMat channel(shrunk, cv::Rect(0, fh * i, fw, fh ));
GpuMat sum(hogluv, cv::Rect(0, (fh + 1) * i, fw + 1, fh + 1));
cv::gpu::integralBuffered(channel, sum, integralBuffer);
}
}
public:
// scales range
float minScale;
float maxScale;
int origObjWidth;
int origObjHeight;
const int shrinkage;
int downscales;
// preallocated buffer 640x480x10 for hogluv + 640x480 got gray
GpuMat plane;
// preallocated buffer for floating point operations
GpuMat fplane;
// temporial mat for cvtColor
GpuMat luv;
// 160x120x10
GpuMat shrunk;
// temporial mat for integrall
GpuMat integralBuffer;
// 161x121x10
GpuMat hogluv;
GpuMat detCounter;
// Cascade from xml
GpuMat octaves;
GpuMat stages;
GpuMat nodes;
GpuMat leaves;
GpuMat levels;
device::icf::CascadeInvoker<device::icf::CascadePolicy> invoker;
enum { BOOST = 0 };
enum
{
FRAME_WIDTH = 640,
FRAME_HEIGHT = 480,
TOTAL_SCALES = 55,
ORIG_OBJECT_WIDTH = 64,
ORIG_OBJECT_HEIGHT = 128,
HOG_BINS = 6,
LUV_BINS = 3,
HOG_LUV_BINS = 10
};
};
cv::gpu::SoftCascade::SoftCascade() : filds(0) {}
......@@ -513,21 +503,15 @@ cv::gpu::SoftCascade::~SoftCascade()
bool cv::gpu::SoftCascade::load( const string& filename, const float minScale, const float maxScale)
{
if (filds)
delete filds;
filds = 0;
if (filds) delete filds;
cv::FileStorage fs(filename, FileStorage::READ);
if (!fs.isOpened()) return false;
filds = new Filds;
Filds& flds = *filds;
if (!flds.fill(fs.getFirstTopLevelNode(), minScale, maxScale)) return false;
return true;
filds = Filds::parseCascade(fs.getFirstTopLevelNode(), minScale, maxScale);
return filds != 0;
}
//================================== synchronous version ============================================================//
void cv::gpu::SoftCascade::detectMultiScale(const GpuMat& colored, const GpuMat& rois,
GpuMat& objects, const int /*rejectfactor*/, int specificScale) const
{
......@@ -562,7 +546,7 @@ void cv::gpu::SoftCascade::detectMultiScale(const GpuMat&, const GpuMat&, GpuMat
cv::Size cv::gpu::SoftCascade::getRoiSize() const
{
return cv::Size(Filds::FRAME_WIDTH / 4, Filds::FRAME_HEIGHT / 4);
return cv::Size(Filds::FRAME_WIDTH / (*filds).shrinkage, Filds::FRAME_HEIGHT / (*filds).shrinkage);
}
#endif
\ No newline at end of file
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment