Skip to content

O
opencv
Commit b83d4add authored by marina.kolpakova's avatar marina.kolpakova
Browse files
Options

memory optimization

parent 4d9c7c10
Show whitespace changes
Inline Side-by-side
Showing with 360 additions and 333 deletions
modules/gpu/src/cuda/isf-sc.cu
View file @ b83d4add
...@@ -41,9 +41,9 @@ ...@@ -41,9 +41,9 @@
//M*/ //M*/
#include <opencv2/gpu/device/common.hpp> #include <opencv2/gpu/device/common.hpp>
// #include <icf.hpp> #include <icf.hpp>
// #include <opencv2/gpu/device/saturate_cast.hpp> // #include <opencv2/gpu/device/saturate_cast.hpp>
// #include <stdio.h> #include <stdio.h>
// #include <float.h> // #include <float.h>
// //#define LOG_CUDA_CASCADE // //#define LOG_CUDA_CASCADE
...@@ -93,6 +93,58 @@ namespace icf { ...@@ -93,6 +93,58 @@ namespace icf {
cudaSafeCall( cudaGetLastError() ); cudaSafeCall( cudaGetLastError() );
cudaSafeCall( cudaDeviceSynchronize() ); cudaSafeCall( cudaDeviceSynchronize() );
} }
texture<float2, cudaTextureType1D, cudaReadModeElementType> tnode;
__global__ void test_kernel(const Level* levels, const Octave* octaves, const float* stages,
const Node* nodes,
PtrStepSz<uchar4> objects)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
Level level = levels[blockIdx.z];
if(x >= level.workRect.x || y >= level.workRect.y) return;
Octave octave = octaves[level.octave];
int st = octave.index * octave.stages;
const int stEnd = st + 1000;//octave.stages;
float confidence = 0.f;
#pragma unroll 8
for(; st < stEnd; ++st)
{
const int nId = st * 3;
const Node node = nodes[nId];
const float stage = stages[st];
confidence += node.rect.x * stage;
}
uchar4 val;
val.x = (int)confidence;
if (x == y) objects(0, threadIdx.x) = val;
}
void detect(const PtrStepSzb& levels, const PtrStepSzb& octaves, const PtrStepSzf& stages,
const PtrStepSzb& nodes, const PtrStepSzb& features,
PtrStepSz<uchar4> objects)
{
int fw = 160;
int fh = 120;
dim3 block(32, 8);
dim3 grid(fw / 32, fh / 8, 47);
const Level* l = (const Level*)levels.ptr();
const Octave* oct = ((const Octave*)octaves.ptr());
const float* st = (const float*)stages.ptr();
const Node* nd = (const Node*)nodes.ptr();
// cudaSafeCall( cudaBindTexture(0, tnode, nodes.data, rgb.cols / size) );
test_kernel<<<grid, block>>>(l, oct, st, nd, objects);
cudaSafeCall( cudaGetLastError());
cudaSafeCall( cudaDeviceSynchronize());
}
} }
}}} }}}
......
modules/gpu/src/icf.hpp
View file @ b83d4add
/*M/////////////////////////////////////////////////////////////////////////////////////// //M///////////////////////////////////////////////////////////////////////////////////////
// //
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
// //
...@@ -38,12 +38,12 @@ ...@@ -38,12 +38,12 @@
// or tort (including negligence or otherwise) arising in any way out of // or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage. // the use of this software, even if advised of the possibility of such damage.
// //
//M*/ //M
// #include <opencv2/gpu/device/common.hpp> #include <opencv2/gpu/device/common.hpp>
// #ifndef __OPENCV_ICF_HPP__ #ifndef __OPENCV_ICF_HPP__
// #define __OPENCV_ICF_HPP__ #define __OPENCV_ICF_HPP__
// #if defined __CUDACC__ // #if defined __CUDACC__
// # define __device __device__ __forceinline__ // # define __device __device__ __forceinline__
...@@ -52,49 +52,62 @@ ...@@ -52,49 +52,62 @@
// #endif // #endif
// namespace cv { namespace gpu { namespace icf { namespace cv { namespace gpu { namespace device {
namespace icf {
// using cv::gpu::PtrStepSzb;
// using cv::gpu::PtrStepSzf; struct __align__(16) Octave
{
// typedef unsigned char uchar; ushort index;
ushort stages;
// struct __align__(16) Octave ushort shrinkage;
// { ushort2 size;
// ushort index; float scale;
// ushort stages;
// ushort shrinkage; Octave(const ushort i, const ushort s, const ushort sh, const ushort2 sz, const float sc)
// ushort2 size; : index(i), stages(s), shrinkage(sh), size(sz), scale(sc) {}
// float scale; };
// Octave(const ushort i, const ushort s, const ushort sh, const ushort2 sz, const float sc) struct __align__(8) Level //is actually 24 bytes
// : index(i), stages(s), shrinkage(sh), size(sz), scale(sc) {} {
// }; int octave;
// struct __align__(8) Level //is actually 24 bytes float relScale;
// { float shrScale; // used for marking detection
// int octave; float scaling[2]; // calculated according to Dollal paper
// // float origScale; //not actually used // for 640x480 we can not get overflow
// float relScale; uchar2 workRect;
// float shrScale; // used for marking detection uchar2 objSize;
// float scaling[2]; // calculated according to Dollal paper
Level(int idx, const Octave& oct, const float scale, const int w, const int h)
// // for 640x480 we can not get overflow : octave(idx), relScale(scale / oct.scale), shrScale (relScale / (float)oct.shrinkage)
// uchar2 workRect; {
// uchar2 objSize; workRect.x = round(w / (float)oct.shrinkage);
workRect.y = round(h / (float)oct.shrinkage);
// Level(int idx, const Octave& oct, const float scale, const int w, const int h)
// : octave(idx), relScale(scale / oct.scale), shrScale (relScale / (float)oct.shrinkage) objSize.x = round(oct.size.x * relScale);
// { objSize.y = round(oct.size.y * relScale);
// workRect.x = round(w / (float)oct.shrinkage); }
// workRect.y = round(h / (float)oct.shrinkage); };
// objSize.x = round(oct.size.x * relScale); struct __align__(8) Node
// objSize.y = round(oct.size.y * relScale); {
// } // int feature;
// }; uchar4 rect;
float threshold;
Node(const uchar4 c, const int t) : rect(c), threshold(t) {}
};
struct __align__(8) Feature
{
int channel;
uchar4 rect;
Feature(const int c, const uchar4 r) : channel(c), rect(r) {}
};
}
}}}
// struct Cascade // struct Cascade
// { // {
// Cascade() {} // Cascade() {}
...@@ -146,21 +159,6 @@ ...@@ -146,21 +159,6 @@
// static const float magnitudeScaling = 1.f ;// / sqrt(2); // static const float magnitudeScaling = 1.f ;// / sqrt(2);
// }; // };
// struct __align__(8) Node
// {
// int feature;
// float threshold;
// Node(const int f, const float t) : feature(f), threshold(t) {}
// };
// struct __align__(8) Feature
// {
// int channel;
// uchar4 rect;
// Feature(const int c, const uchar4 r) : channel(c), rect(r) {}
// };
// }}} // }}}
// #endif #endif
\ No newline at end of file \ No newline at end of file
modules/gpu/src/softcascade.cpp
View file @ b83d4add
...@@ -53,12 +53,15 @@ void cv::gpu::SoftCascade::detectMultiScale(const GpuMat&, const GpuMat&, GpuMat ...@@ -53,12 +53,15 @@ void cv::gpu::SoftCascade::detectMultiScale(const GpuMat&, const GpuMat&, GpuMat
#else #else
// #include <icf.hpp> #include <icf.hpp>
namespace cv { namespace gpu { namespace device { namespace cv { namespace gpu { namespace device {
namespace icf { namespace icf {
void fillBins(cv::gpu::PtrStepSzb hogluv, const cv::gpu::PtrStepSzf& nangle, void fillBins(cv::gpu::PtrStepSzb hogluv, const cv::gpu::PtrStepSzf& nangle,
const int fw, const int fh, const int bins); const int fw, const int fh, const int bins);
void detect(const PtrStepSzb& levels, const PtrStepSzb& octaves, const PtrStepSzf& stages,
const PtrStepSzb& nodes, const PtrStepSzb& features,
PtrStepSz<uchar4> objects);
} }
}}} }}}
...@@ -82,19 +85,20 @@ struct cv::gpu::SoftCascade::Filds ...@@ -82,19 +85,20 @@ struct cv::gpu::SoftCascade::Filds
integralBuffer.create(shrunk.rows + 1 * HOG_LUV_BINS, shrunk.cols + 1, CV_32SC1); integralBuffer.create(shrunk.rows + 1 * HOG_LUV_BINS, shrunk.cols + 1, CV_32SC1);
hogluv.create((FRAME_HEIGHT / 4 + 1) * HOG_LUV_BINS, FRAME_WIDTH / 4 + 1, CV_32SC1); hogluv.create((FRAME_HEIGHT / 4 + 1) * HOG_LUV_BINS, FRAME_WIDTH / 4 + 1, CV_32SC1);
} }
// // scales range
// float minScale;
// float maxScale;
// int origObjWidth; // scales range
// int origObjHeight; float minScale;
float maxScale;
// GpuMat octaves; int origObjWidth;
// GpuMat stages; int origObjHeight;
// GpuMat nodes;
// GpuMat leaves; GpuMat octaves;
// GpuMat features; GpuMat stages;
// GpuMat levels; GpuMat nodes;
GpuMat leaves;
GpuMat features;
GpuMat levels;
// preallocated buffer 640x480x10 for hogluv + 640x480 got gray // preallocated buffer 640x480x10 for hogluv + 640x480 got gray
GpuMat plane; GpuMat plane;
...@@ -114,312 +118,285 @@ struct cv::gpu::SoftCascade::Filds ...@@ -114,312 +118,285 @@ struct cv::gpu::SoftCascade::Filds
// 161x121x10 // 161x121x10
GpuMat hogluv; GpuMat hogluv;
// // will be removed in final version std::vector<float> scales;
// // temp matrix for sobel and cartToPolar
// GpuMat dfdx, dfdy, angle, mag, nmag, nangle;
// std::vector<float> scales;
// icf::Cascade cascade;
// icf::ChannelStorage storage;
// enum { BOOST = 0 }; enum { BOOST = 0 };
enum enum
{ {
FRAME_WIDTH = 640, FRAME_WIDTH = 640,
FRAME_HEIGHT = 480, FRAME_HEIGHT = 480,
// TOTAL_SCALES = 55, TOTAL_SCALES = 55,
// CLASSIFIERS = 5, // CLASSIFIERS = 5,
// ORIG_OBJECT_WIDTH = 64, ORIG_OBJECT_WIDTH = 64,
// ORIG_OBJECT_HEIGHT = 128, ORIG_OBJECT_HEIGHT = 128,
HOG_BINS = 6, HOG_BINS = 6,
LUV_BINS = 3, LUV_BINS = 3,
HOG_LUV_BINS = 10 HOG_LUV_BINS = 10
}; };
// bool fill(const FileNode &root, const float mins, const float maxs); bool fill(const FileNode &root, const float mins, const float maxs);
// void detect(cv::gpu::GpuMat objects, cudaStream_t stream) const void detect(cv::gpu::GpuMat objects, cudaStream_t stream) const
// { {
// cascade.detect(hogluv, objects, stream); device::icf::detect(levels, octaves, stages, nodes, features, objects);
// } }
// private: private:
// void calcLevels(const std::vector<icf::Octave>& octs, void calcLevels(const std::vector<device::icf::Octave>& octs,
// int frameW, int frameH, int nscales); int frameW, int frameH, int nscales);
// typedef std::vector<icf::Octave>::const_iterator octIt_t; typedef std::vector<device::icf::Octave>::const_iterator octIt_t;
// int fitOctave(const std::vector<icf::Octave>& octs, const float& logFactor) const int fitOctave(const std::vector<device::icf::Octave>& octs, const float& logFactor) const
// { {
// float minAbsLog = FLT_MAX; float minAbsLog = FLT_MAX;
// int res = 0; int res = 0;
// for (int oct = 0; oct < (int)octs.size(); ++oct) for (int oct = 0; oct < (int)octs.size(); ++oct)
// { {
// const icf::Octave& octave =octs[oct]; const device::icf::Octave& octave =octs[oct];
// float logOctave = ::log(octave.scale); float logOctave = ::log(octave.scale);
// float logAbsScale = ::fabs(logFactor - logOctave); float logAbsScale = ::fabs(logFactor - logOctave);
// if(logAbsScale < minAbsLog) if(logAbsScale < minAbsLog)
// { {
// res = oct; res = oct;
// minAbsLog = logAbsScale; minAbsLog = logAbsScale;
// } }
// } }
// return res; return res;
// } }
}; };
// inline bool cv::gpu::SoftCascade::Filds::fill(const FileNode &root, const float mins, const float maxs) inline bool cv::gpu::SoftCascade::Filds::fill(const FileNode &root, const float mins, const float maxs)
// { {
// minScale = mins; using namespace device::icf;
// maxScale = maxs; 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"; // cascade properties
// static const char *const SC_ORIG_H = "height"; static const char *const SC_STAGE_TYPE = "stageType";
static const char *const SC_BOOST = "BOOST";
// static const char *const SC_OCTAVES = "octaves"; static const char *const SC_FEATURE_TYPE = "featureType";
// static const char *const SC_STAGES = "stages"; static const char *const SC_ICF = "ICF";
// static const char *const SC_FEATURES = "features";
// static const char *const SC_WEEK = "weakClassifiers"; static const char *const SC_ORIG_W = "width";
// static const char *const SC_INTERNAL = "internalNodes"; static const char *const SC_ORIG_H = "height";
// static const char *const SC_LEAF = "leafValues";
// static const char *const SC_OCT_SCALE = "scale"; static const char *const SC_OCTAVES = "octaves";
// static const char *const SC_OCT_STAGES = "stageNum"; static const char *const SC_STAGES = "stages";
// static const char *const SC_OCT_SHRINKAGE = "shrinkingFactor"; static const char *const SC_FEATURES = "features";
// static const char *const SC_STAGE_THRESHOLD = "stageThreshold"; 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_F_CHANNEL = "channel"; static const char *const SC_OCT_SCALE = "scale";
// static const char * const SC_F_RECT = "rect"; static const char *const SC_OCT_STAGES = "stageNum";
static const char *const SC_OCT_SHRINKAGE = "shrinkingFactor";
// // only Ada Boost supported static const char *const SC_STAGE_THRESHOLD = "stageThreshold";
// std::string stageTypeStr = (string)root[SC_STAGE_TYPE];
// CV_Assert(stageTypeStr == SC_BOOST);
// // only HOG-like integral channel features cupported static const char * const SC_F_CHANNEL = "channel";
// string featureTypeStr = (string)root[SC_FEATURE_TYPE]; static const char * const SC_F_RECT = "rect";
// CV_Assert(featureTypeStr == SC_ICF);
// origObjWidth = (int)root[SC_ORIG_W]; // only Ada Boost supported
// CV_Assert(origObjWidth == ORIG_OBJECT_WIDTH); std::string stageTypeStr = (string)root[SC_STAGE_TYPE];
CV_Assert(stageTypeStr == SC_BOOST);
// origObjHeight = (int)root[SC_ORIG_H]; // only HOG-like integral channel features cupported
// CV_Assert(origObjHeight == ORIG_OBJECT_HEIGHT); string featureTypeStr = (string)root[SC_FEATURE_TYPE];
CV_Assert(featureTypeStr == SC_ICF);
// FileNode fn = root[SC_OCTAVES]; origObjWidth = (int)root[SC_ORIG_W];
// if (fn.empty()) return false; CV_Assert(origObjWidth == ORIG_OBJECT_WIDTH);
// std::vector<icf::Octave> voctaves; origObjHeight = (int)root[SC_ORIG_H];
// std::vector<float> vstages; CV_Assert(origObjHeight == ORIG_OBJECT_HEIGHT);
// std::vector<icf::Node> vnodes;
// std::vector<float> vleaves;
// std::vector<icf::Feature> vfeatures;
// scales.clear();
// // std::vector<Level> levels; FileNode fn = root[SC_OCTAVES];
if (fn.empty()) return false;
// FileNodeIterator it = fn.begin(), it_end = fn.end(); std::vector<Octave> voctaves;
// int feature_offset = 0; std::vector<float> vstages;
// ushort octIndex = 0; std::vector<Node> vnodes;
// ushort shrinkage = 1; std::vector<float> vleaves;
std::vector<Feature> vfeatures;
scales.clear();
// for (; it != it_end; ++it) FileNodeIterator it = fn.begin(), it_end = fn.end();
// { int feature_offset = 0;
// FileNode fns = *it; ushort octIndex = 0;
// float scale = (float)fns[SC_OCT_SCALE]; ushort shrinkage = 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]);
// icf::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;
// 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]; for (; it != it_end; ++it)
// FileNodeIterator ftr = fns.begin(), ft_end = fns.end(); {
// for (; ftr != ft_end; ++ftr) FileNode fns = *it;
// { float scale = (float)fns[SC_OCT_SCALE];
// fns = (*ftr)[SC_INTERNAL]; scales.push_back(scale);
// FileNodeIterator inIt = fns.begin(), inIt_end = fns.end(); ushort nstages = saturate_cast<ushort>((int)fns[SC_OCT_STAGES]);
// for (; inIt != inIt_end;) ushort2 size;
// { size.x = cvRound(ORIG_OBJECT_WIDTH * scale);
// int feature = (int)(*(inIt +=2)++) + feature_offset; size.y = cvRound(ORIG_OBJECT_HEIGHT * scale);
// float th = (float)(*(inIt++)); shrinkage = saturate_cast<ushort>((int)fns[SC_OCT_SHRINKAGE]);
// vnodes.push_back(icf::Node(feature, th));
// } Octave octave(octIndex, nstages, shrinkage, size, scale);
CV_Assert(octave.stages > 0);
// fns = (*ftr)[SC_LEAF]; voctaves.push_back(octave);
// inIt = fns.begin(), inIt_end = fns.end();
// for (; inIt != inIt_end; ++inIt) FileNode ffs = fns[SC_FEATURES];
// vleaves.push_back((float)(*inIt)); if (ffs.empty()) return false;
// }
// } fns = fns[SC_STAGES];
if (fn.empty()) return false;
// st = ffs.begin(), st_end = ffs.end();
// for (; st != st_end; ++st ) // for each stage (~ decision tree with H = 2)
// { FileNodeIterator st = fns.begin(), st_end = fns.end();
// cv::FileNode rn = (*st)[SC_F_RECT]; for (; st != st_end; ++st )
// cv::FileNodeIterator r_it = rn.begin(); {
// uchar4 rect; fns = *st;
// rect.x = saturate_cast<uchar>((int)*(r_it++)); vstages.push_back((float)fns[SC_STAGE_THRESHOLD]);
// rect.y = saturate_cast<uchar>((int)*(r_it++));
// rect.z = saturate_cast<uchar>((int)*(r_it++));
// rect.w = saturate_cast<uchar>((int)*(r_it++));
// vfeatures.push_back(icf::Feature((int)(*st)[SC_F_CHANNEL], rect));
// }
// feature_offset += octave.stages * 3; fns = fns[SC_WEEK];
// ++octIndex; 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;
float th = (float)(*(inIt++));
uchar4 rect;
vnodes.push_back(Node(rect, th));
}
// // upload in gpu memory fns = (*ftr)[SC_LEAF];
// octaves.upload(cv::Mat(1, voctaves.size() * sizeof(icf::Octave), CV_8UC1, (uchar*)&(voctaves[0]) )); inIt = fns.begin(), inIt_end = fns.end();
// CV_Assert(!octaves.empty()); for (; inIt != inIt_end; ++inIt)
vleaves.push_back((float)(*inIt));
}
}
// stages.upload(cv::Mat(vstages).reshape(1,1)); st = ffs.begin(), st_end = ffs.end();
// CV_Assert(!stages.empty()); for (; st != st_end; ++st )
{
cv::FileNode rn = (*st)[SC_F_RECT];
cv::FileNodeIterator r_it = rn.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++));
vfeatures.push_back(Feature((int)(*st)[SC_F_CHANNEL], rect));
}
// nodes.upload(cv::Mat(1, vnodes.size() * sizeof(icf::Node), CV_8UC1, (uchar*)&(vnodes[0]) )); feature_offset += octave.stages * 3;
// CV_Assert(!nodes.empty()); ++octIndex;
}
// leaves.upload(cv::Mat(vleaves).reshape(1,1)); // upload in gpu memory
// CV_Assert(!leaves.empty()); octaves.upload(cv::Mat(1, voctaves.size() * sizeof(Octave), CV_8UC1, (uchar*)&(voctaves[0]) ));
CV_Assert(!octaves.empty());
// features.upload(cv::Mat(1, vfeatures.size() * sizeof(icf::Feature), CV_8UC1, (uchar*)&(vfeatures[0]) )); stages.upload(cv::Mat(vstages).reshape(1,1));
// CV_Assert(!features.empty()); CV_Assert(!stages.empty());
// // compute levels nodes.upload(cv::Mat(1, vnodes.size() * sizeof(Node), CV_8UC1, (uchar*)&(vnodes[0]) ));
// calcLevels(voctaves, FRAME_WIDTH, FRAME_HEIGHT, TOTAL_SCALES); CV_Assert(!nodes.empty());
// CV_Assert(!levels.empty());
// //init Cascade leaves.upload(cv::Mat(vleaves).reshape(1,1));
// cascade = icf::Cascade(octaves, stages, nodes, leaves, features, levels); CV_Assert(!leaves.empty());
// // allocate buffers features.upload(cv::Mat(1, vfeatures.size() * sizeof(Feature), CV_8UC1, (uchar*)&(vfeatures[0]) ));
// dmem.create(FRAME_HEIGHT * (HOG_LUV_BINS + 1), FRAME_WIDTH, CV_8UC1); CV_Assert(!features.empty());
// shrunk.create(FRAME_HEIGHT / shrinkage * HOG_LUV_BINS, FRAME_WIDTH / shrinkage, CV_8UC1);
// // hogluv.create( (FRAME_HEIGHT / shrinkage + 1) * HOG_LUV_BINS, (FRAME_WIDTH / shrinkage + 1), CV_16UC1);
// hogluv.create( (FRAME_HEIGHT / shrinkage + 1) * HOG_LUV_BINS, (FRAME_WIDTH / shrinkage + 1), CV_32SC1);
// luv.create(FRAME_HEIGHT, FRAME_WIDTH, CV_8UC3);
// integralBuffer.create(shrunk.rows + 1 * HOG_LUV_BINS, shrunk.cols + 1, CV_32SC1);
// dfdx.create(FRAME_HEIGHT, FRAME_WIDTH, CV_32FC1); // compute levels
// dfdy.create(FRAME_HEIGHT, FRAME_WIDTH, CV_32FC1); calcLevels(voctaves, FRAME_WIDTH, FRAME_HEIGHT, TOTAL_SCALES);
// angle.create(FRAME_HEIGHT, FRAME_WIDTH, CV_32FC1); CV_Assert(!levels.empty());
// mag.create(FRAME_HEIGHT, FRAME_WIDTH, CV_32FC1);
// nmag.create(FRAME_HEIGHT, FRAME_WIDTH, CV_32FC1); return true;
// nangle.create(FRAME_HEIGHT, FRAME_WIDTH, CV_32FC1); }
// storage = icf::ChannelStorage(dmem, shrunk, hogluv, shrinkage); namespace {
// return true; struct CascadeIntrinsics
// } {
static const float lambda = 1.099f, a = 0.89f;
// namespace { static float getFor(int channel, float scaling)
// struct CascadeIntrinsics {
// { CV_Assert(channel < 10);
// static const float lambda = 1.099f, a = 0.89f;
// static float getFor(int channel, float scaling) if (fabs(scaling - 1.f) < FLT_EPSILON)
// { return 1.f;
// CV_Assert(channel < 10);
// if (fabs(scaling - 1.f) < FLT_EPSILON) // according to R. Benenson, M. Mathias, R. Timofte and L. Van Gool's and Dallal's papers
// return 1.f; static const float A[2][2] =
{ //channel <= 6, otherwise
{ 0.89f, 1.f}, // down
{ 1.00f, 1.f} // up
};
// // according to R. Benenson, M. Mathias, R. Timofte and L. Van Gool's and Dallal's papers static const float B[2][2] =
// static const float A[2][2] = { //channel <= 6, otherwise
// { //channel <= 6, otherwise { 1.099f / log(2), 2.f}, // down
// { 0.89f, 1.f}, // down { 0.f, 2.f} // up
// { 1.00f, 1.f} // up };
// };
// static const float B[2][2] = float a = A[(int)(scaling >= 1)][(int)(channel > 6)];
// { //channel <= 6, otherwise float b = B[(int)(scaling >= 1)][(int)(channel > 6)];
// { 1.099f / log(2), 2.f}, // down
// { 0.f, 2.f} // up
// };
// float a = A[(int)(scaling >= 1)][(int)(channel > 6)]; // printf("!!! scaling: %f %f %f -> %f\n", scaling, a, b, a * pow(scaling, b));
// float b = B[(int)(scaling >= 1)][(int)(channel > 6)]; return a * pow(scaling, b);
}
};
}
// // printf("!!! scaling: %f %f %f -> %f\n", scaling, a, b, a * pow(scaling, b)); inline void cv::gpu::SoftCascade::Filds::calcLevels(const std::vector<device::icf::Octave>& octs,
// return a * pow(scaling, b); int frameW, int frameH, int nscales)
// } {
// }; CV_Assert(nscales > 1);
// } using device::icf::Level;
// inline void cv::gpu::SoftCascade::Filds::calcLevels(const std::vector<icf::Octave>& octs, std::vector<Level> vlevels;
// int frameW, int frameH, int nscales) float logFactor = (::log(maxScale) - ::log(minScale)) / (nscales -1);
// {
// CV_Assert(nscales > 1);
// std::vector<icf::Level> vlevels; float scale = minScale;
// float logFactor = (::log(maxScale) - ::log(minScale)) / (nscales -1); 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 (::fabs(scale - maxScale) < FLT_EPSILON) break;
scale = ::std::min(maxScale, ::expf(::log(scale) + logFactor));
// printf("level: %d (%f %f) [%f %f] (%d %d) (%d %d)\n", level.octave, level.relScale, level.shrScale,
// level.scaling[0], level.scaling[1], level.workRect.x, level.workRect.y, level.objSize.x,
//level.objSize.y);
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;
}
// float scale = minScale; levels.upload(cv::Mat(1, vlevels.size() * sizeof(Level), CV_8UC1, (uchar*)&(vlevels[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);
// icf::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 (::fabs(scale - maxScale) < FLT_EPSILON) break;
// scale = ::std::min(maxScale, ::expf(::log(scale) + logFactor));
// // printf("level: %d (%f %f) [%f %f] (%d %d) (%d %d)\n", level.octave, level.relScale, level.shrScale,
// // level.scaling[0], level.scaling[1], level.workRect.x, level.workRect.y, level.objSize.x,
//level.objSize.y);
// // 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(icf::Level), CV_8UC1, (uchar*)&(vlevels[0]) ));
// }
cv::gpu::SoftCascade::SoftCascade() : filds(0) {} cv::gpu::SoftCascade::SoftCascade() : filds(0) {}
...@@ -444,7 +421,7 @@ bool cv::gpu::SoftCascade::load( const string& filename, const float minScale, c ...@@ -444,7 +421,7 @@ bool cv::gpu::SoftCascade::load( const string& filename, const float minScale, c
filds = new Filds; filds = new Filds;
Filds& flds = *filds; Filds& flds = *filds;
// if (!flds.fill(fs.getFirstTopLevelNode(), minScale, maxScale)) return false; if (!flds.fill(fs.getFirstTopLevelNode(), minScale, maxScale)) return false;
return true; return true;
} }
...@@ -538,7 +515,7 @@ void cv::gpu::SoftCascade::detectMultiScale(const GpuMat& colored, const GpuMat& ...@@ -538,7 +515,7 @@ void cv::gpu::SoftCascade::detectMultiScale(const GpuMat& colored, const GpuMat&
cudaStream_t stream = StreamAccessor::getStream(s); cudaStream_t stream = StreamAccessor::getStream(s);
// detection // detection
// flds.detect(objects, stream); flds.detect(objects, stream);
// // flds.storage.frame(colored, stream); // // flds.storage.frame(colored, stream);
} }
......
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