mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2024-11-26 12:10:49 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

add factory method for Fields structure

Browse Source
This commit is contained in:
marina.kolpakova 2012-10-11 19:11:39 +04:00
parent 0898c3c651
commit f196e9fda4
1 changed files with 281 additions and 297 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

608
modules/gpu/src/softcascade.cpp
View File

@ -78,77 +78,255 @@ namespace imgproc
struct cv::gpu::SoftCascade::Filds
{
struct CascadeIntrinsics
{
static const float lambda = 1.099f, a = 0.89f;
Filds()
static float getFor(int channel, float scaling)
{
CV_Assert(channel < 10);
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);
}
};
static Filds* parseCascade(const FileNode &root, const float mins, const float maxs)
{
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";
static const char *const SC_WEEK = "weakClassifiers";
static const char *const SC_INTERNAL = "internalNodes";
static const char *const SC_LEAF = "leafValues";
static const char *const SC_OCT_SCALE = "scale";
static const char *const SC_OCT_STAGES = "stageNum";
static const char *const SC_OCT_SHRINKAGE = "shrinkingFactor";
static const char *const SC_STAGE_THRESHOLD = "stageThreshold";
static const char * const SC_F_CHANNEL = "channel";
static const char * const SC_F_RECT = "rect";
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;
FileNodeIterator it = fn.begin(), it_end = fn.end();
int feature_offset = 0;
ushort octIndex = 0;
ushort shrinkage = 1;
for (; it != it_end; ++it)
{
FileNode fns = *it;
float scale = (float)fns[SC_OCT_SCALE];
bool isUPOctave = scale >= 1;
ushort nstages = saturate_cast<ushort>((int)fns[SC_OCT_STAGES]);
ushort2 size;
size.x = cvRound(ORIG_OBJECT_WIDTH * scale);
size.y = cvRound(ORIG_OBJECT_HEIGHT * scale);
shrinkage = saturate_cast<ushort>((int)fns[SC_OCT_SHRINKAGE]);
Octave octave(octIndex, nstages, shrinkage, size, scale);
CV_Assert(octave.stages > 0);
voctaves.push_back(octave);
FileNode ffs = fns[SC_FEATURES];
if (ffs.empty()) return false;
FileNodeIterator ftrs = ffs.begin();
fns = fns[SC_STAGES];
if (fn.empty()) return false;
// for each stage (~ decision tree with H = 2)
FileNodeIterator st = fns.begin(), st_end = fns.end();
for (; st != st_end; ++st )
{
fns = *st;
vstages.push_back((float)fns[SC_STAGE_THRESHOLD]);
fns = fns[SC_WEEK];
FileNodeIterator ftr = fns.begin(), ft_end = fns.end();
for (; ftr != ft_end; ++ftr)
{
fns = (*ftr)[SC_INTERNAL];
FileNodeIterator inIt = fns.begin(), inIt_end = fns.end();
for (; inIt != inIt_end;)
{
// int feature = (int)(*(inIt +=2)) + feature_offset;
inIt +=3;
// extract feature, Todo:check it
uint th = saturate_cast<uint>((float)(*(inIt++)));
cv::FileNode ftn = (*ftrs)[SC_F_RECT];
cv::FileNodeIterator r_it = ftn.begin();
uchar4 rect;
rect.x = saturate_cast<uchar>((int)*(r_it++));
rect.y = saturate_cast<uchar>((int)*(r_it++));
rect.z = saturate_cast<uchar>((int)*(r_it++));
rect.w = saturate_cast<uchar>((int)*(r_it++));
if (isUPOctave)
{
rect.z -= rect.x;
rect.w -= rect.y;
}
uint channel = saturate_cast<uint>((int)(*ftrs)[SC_F_CHANNEL]);
vnodes.push_back(Node(rect, channel, th));
++ftrs;
}
fns = (*ftr)[SC_LEAF];
inIt = fns.begin(), inIt_end = fns.end();
for (; inIt != inIt_end; ++inIt)
vleaves.push_back((float)(*inIt));
}
}
feature_offset += octave.stages * 3;
++octIndex;
}
cv::Mat hoctaves(1, voctaves.size() * sizeof(Octave), CV_8UC1, (uchar*)&(voctaves[0]));
CV_Assert(!hoctaves.empty());
cv::Mat hstages(cv::Mat(vstages).reshape(1,1));
CV_Assert(!hstages.empty());
cv::Mat hnodes(1, vnodes.size() * sizeof(Node), CV_8UC1, (uchar*)&(vnodes[0]) );
CV_Assert(!hnodes.empty());
cv::Mat hleaves(cv::Mat(vleaves).reshape(1,1));
CV_Assert(!hleaves.empty());
std::vector<Level> vlevels;
float logFactor = (::log(maxs) - ::log(mins)) / (TOTAL_SCALES -1);
float scale = mins;
int downscales = 0;
for (int sc = 0; sc < TOTAL_SCALES; ++sc)
{
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(voctaves, logScale);
Level level(fit, voctaves[fit], scale, width, height);
level.scaling[0] = CascadeIntrinsics::getFor(0, level.relScale);
level.scaling[1] = CascadeIntrinsics::getFor(9, level.relScale);
if (!width || !height)
break;
else
{
vlevels.push_back(level);
if (voctaves[fit].scale < 1) ++downscales;
}
if (::fabs(scale - maxs) < FLT_EPSILON) break;
scale = ::std::min(maxs, ::expf(::log(scale) + logFactor));
// std::cout << "level " << sc
// << " octeve "
// << vlevels[sc].octave
// << " relScale "
// << vlevels[sc].relScale
// << " " << vlevels[sc].shrScale
// << " [" << (int)vlevels[sc].objSize.x
// << " " << (int)vlevels[sc].objSize.y << "] ["
// << (int)vlevels[sc].workRect.x << " " << (int)vlevels[sc].workRect.y << "]" << std::endl;
}
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 / 4 * HOG_LUV_BINS, FRAME_WIDTH / 4, CV_8UC1);
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 / 4 + 1) * HOG_LUV_BINS, FRAME_WIDTH / 4 + 64, 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);
}
// 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);
}
@ -169,11 +347,9 @@ struct cv::gpu::SoftCascade::Filds
}
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
static int fitOctave(const std::vector<device::icf::Octave>& octs, const float& logFactor)
{
float minAbsLog = FLT_MAX;
int res = 0;
@ -257,248 +433,62 @@ private:
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
};
};
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";
static const char *const SC_ORIG_W = "width";
static const char *const SC_ORIG_H = "height";
static const char *const SC_OCTAVES = "octaves";
static const char *const SC_STAGES = "stages";
static const char *const SC_FEATURES = "features";
static const char *const SC_WEEK = "weakClassifiers";
static const char *const SC_INTERNAL = "internalNodes";
static const char *const SC_LEAF = "leafValues";
static const char *const SC_OCT_SCALE = "scale";
static const char *const SC_OCT_STAGES = "stageNum";
static const char *const SC_OCT_SHRINKAGE = "shrinkingFactor";
static const char *const SC_STAGE_THRESHOLD = "stageThreshold";
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;
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;
ushort octIndex = 0;
ushort shrinkage = 1;
for (; it != it_end; ++it)
{
FileNode fns = *it;
float scale = (float)fns[SC_OCT_SCALE];
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);
size.y = cvRound(ORIG_OBJECT_HEIGHT * scale);
shrinkage = saturate_cast<ushort>((int)fns[SC_OCT_SHRINKAGE]);
Octave octave(octIndex, nstages, shrinkage, size, scale);
CV_Assert(octave.stages > 0);
voctaves.push_back(octave);
FileNode ffs = fns[SC_FEATURES];
if (ffs.empty()) return false;
FileNodeIterator ftrs = ffs.begin();
fns = fns[SC_STAGES];
if (fn.empty()) return false;
// for each stage (~ decision tree with H = 2)
FileNodeIterator st = fns.begin(), st_end = fns.end();
for (; st != st_end; ++st )
{
fns = *st;
vstages.push_back((float)fns[SC_STAGE_THRESHOLD]);
fns = fns[SC_WEEK];
FileNodeIterator ftr = fns.begin(), ft_end = fns.end();
for (; ftr != ft_end; ++ftr)
{
fns = (*ftr)[SC_INTERNAL];
FileNodeIterator inIt = fns.begin(), inIt_end = fns.end();
for (; inIt != inIt_end;)
{
// int feature = (int)(*(inIt +=2)) + feature_offset;
inIt +=3;
// extract feature, Todo:check it
uint th = saturate_cast<uint>((float)(*(inIt++)));
cv::FileNode ftn = (*ftrs)[SC_F_RECT];
cv::FileNodeIterator r_it = ftn.begin();
uchar4 rect;
rect.x = saturate_cast<uchar>((int)*(r_it++));
rect.y = saturate_cast<uchar>((int)*(r_it++));
rect.z = saturate_cast<uchar>((int)*(r_it++));
rect.w = saturate_cast<uchar>((int)*(r_it++));
if (isUPOctave)
{
rect.z -= rect.x;
rect.w -= rect.y;
}
uint channel = saturate_cast<uint>((int)(*ftrs)[SC_F_CHANNEL]);
vnodes.push_back(Node(rect, channel, th));
++ftrs;
}
fns = (*ftr)[SC_LEAF];
inIt = fns.begin(), inIt_end = fns.end();
for (; inIt != inIt_end; ++inIt)
vleaves.push_back((float)(*inIt));
}
}
feature_offset += octave.stages * 3;
++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;
}
namespace {
struct CascadeIntrinsics
{
static const float lambda = 1.099f, a = 0.89f;
static float getFor(int channel, float scaling)
{
CV_Assert(channel < 10);
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;
std::vector<Level> vlevels;
float logFactor = (::log(maxScale) - ::log(minScale)) / (nscales -1);
float scale = minScale;
downscales = 0;
for (int sc = 0; sc < nscales; ++sc)
{
int width = ::std::max(0.0f, frameW - (origObjWidth * scale));
int height = ::std::max(0.0f, frameH - (origObjHeight * scale));
float logScale = ::log(scale);
int fit = fitOctave(octs, logScale);
Level level(fit, octs[fit], scale, width, height);
level.scaling[0] = CascadeIntrinsics::getFor(0, level.relScale);
level.scaling[1] = CascadeIntrinsics::getFor(9, level.relScale);
if (!width || !height)
break;
else
{
vlevels.push_back(level);
if (octs[fit].scale < 1) ++downscales;
}
if (::fabs(scale - maxScale) < FLT_EPSILON) break;
scale = ::std::min(maxScale, ::expf(::log(scale) + logFactor));
// std::cout << "level " << sc
// << " octeve "
// << vlevels[sc].octave
// << " relScale "
// << vlevels[sc].relScale
// << " " << vlevels[sc].shrScale
// << " [" << (int)vlevels[sc].objSize.x
// << " " << (int)vlevels[sc].objSize.y << "] ["
// << (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::gpu::SoftCascade::SoftCascade() : filds(0) {}
cv::gpu::SoftCascade::SoftCascade( const string& filename, const float minScale, const float maxScale) : 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