add part of opencv
This commit is contained in:
Tang1705
1763
Lib/opencv/sources/modules/objdetect/src/cascadedetect.cpp
Normal file
1763
Lib/opencv/sources/modules/objdetect/src/cascadedetect.cpp
Normal file
File diff suppressed because it is too large
Load Diff
656
Lib/opencv/sources/modules/objdetect/src/cascadedetect.hpp
Normal file
656
Lib/opencv/sources/modules/objdetect/src/cascadedetect.hpp
Normal file
@@ -0,0 +1,656 @@
|
||||
#pragma once
|
||||
|
||||
#include "opencv2/core/ocl.hpp"
|
||||
|
||||
namespace cv
|
||||
{
|
||||
|
||||
void clipObjects(Size sz, std::vector<Rect>& objects,
|
||||
std::vector<int>* a, std::vector<double>* b);
|
||||
|
||||
class FeatureEvaluator
|
||||
{
|
||||
public:
|
||||
enum
|
||||
{
|
||||
HAAR = 0,
|
||||
LBP = 1,
|
||||
HOG = 2
|
||||
};
|
||||
|
||||
struct ScaleData
|
||||
{
|
||||
ScaleData() { scale = 0.f; layer_ofs = ystep = 0; }
|
||||
Size getWorkingSize(Size winSize) const
|
||||
{
|
||||
return Size(std::max(szi.width - winSize.width, 0),
|
||||
std::max(szi.height - winSize.height, 0));
|
||||
}
|
||||
|
||||
float scale;
|
||||
Size szi;
|
||||
int layer_ofs, ystep;
|
||||
};
|
||||
|
||||
virtual ~FeatureEvaluator();
|
||||
|
||||
virtual bool read(const FileNode& node, Size origWinSize);
|
||||
virtual Ptr<FeatureEvaluator> clone() const;
|
||||
virtual int getFeatureType() const;
|
||||
int getNumChannels() const { return nchannels; }
|
||||
|
||||
virtual bool setImage(InputArray img, const std::vector<float>& scales);
|
||||
virtual bool setWindow(Point p, int scaleIdx);
|
||||
const ScaleData& getScaleData(int scaleIdx) const
|
||||
{
|
||||
CV_Assert( 0 <= scaleIdx && scaleIdx < (int)scaleData->size());
|
||||
return scaleData->at(scaleIdx);
|
||||
}
|
||||
virtual void getUMats(std::vector<UMat>& bufs);
|
||||
virtual void getMats();
|
||||
|
||||
Size getLocalSize() const { return localSize; }
|
||||
Size getLocalBufSize() const { return lbufSize; }
|
||||
|
||||
virtual float calcOrd(int featureIdx) const;
|
||||
virtual int calcCat(int featureIdx) const;
|
||||
|
||||
static Ptr<FeatureEvaluator> create(int type);
|
||||
|
||||
protected:
|
||||
enum { SBUF_VALID=1, USBUF_VALID=2 };
|
||||
int sbufFlag;
|
||||
|
||||
bool updateScaleData( Size imgsz, const std::vector<float>& _scales );
|
||||
virtual void computeChannels( int, InputArray ) {}
|
||||
virtual void computeOptFeatures() {}
|
||||
|
||||
Size origWinSize, sbufSize, localSize, lbufSize;
|
||||
int nchannels;
|
||||
Mat sbuf, rbuf;
|
||||
UMat urbuf, usbuf, ufbuf, uscaleData;
|
||||
|
||||
Ptr<std::vector<ScaleData> > scaleData;
|
||||
};
|
||||
|
||||
|
||||
class CascadeClassifierImpl CV_FINAL : public BaseCascadeClassifier
|
||||
{
|
||||
public:
|
||||
CascadeClassifierImpl();
|
||||
virtual ~CascadeClassifierImpl() CV_OVERRIDE;
|
||||
|
||||
bool empty() const CV_OVERRIDE;
|
||||
bool load( const String& filename ) CV_OVERRIDE;
|
||||
void read( const FileNode& node ) CV_OVERRIDE;
|
||||
bool read_( const FileNode& node );
|
||||
void detectMultiScale( InputArray image,
|
||||
CV_OUT std::vector<Rect>& objects,
|
||||
double scaleFactor = 1.1,
|
||||
int minNeighbors = 3, int flags = 0,
|
||||
Size minSize = Size(),
|
||||
Size maxSize = Size() ) CV_OVERRIDE;
|
||||
|
||||
void detectMultiScale( InputArray image,
|
||||
CV_OUT std::vector<Rect>& objects,
|
||||
CV_OUT std::vector<int>& numDetections,
|
||||
double scaleFactor=1.1,
|
||||
int minNeighbors=3, int flags=0,
|
||||
Size minSize=Size(),
|
||||
Size maxSize=Size() ) CV_OVERRIDE;
|
||||
|
||||
void detectMultiScale( InputArray image,
|
||||
CV_OUT std::vector<Rect>& objects,
|
||||
CV_OUT std::vector<int>& rejectLevels,
|
||||
CV_OUT std::vector<double>& levelWeights,
|
||||
double scaleFactor = 1.1,
|
||||
int minNeighbors = 3, int flags = 0,
|
||||
Size minSize = Size(),
|
||||
Size maxSize = Size(),
|
||||
bool outputRejectLevels = false ) CV_OVERRIDE;
|
||||
|
||||
|
||||
bool isOldFormatCascade() const CV_OVERRIDE;
|
||||
Size getOriginalWindowSize() const CV_OVERRIDE;
|
||||
int getFeatureType() const CV_OVERRIDE;
|
||||
void* getOldCascade() CV_OVERRIDE;
|
||||
|
||||
void setMaskGenerator(const Ptr<MaskGenerator>& maskGenerator) CV_OVERRIDE;
|
||||
Ptr<MaskGenerator> getMaskGenerator() CV_OVERRIDE;
|
||||
|
||||
protected:
|
||||
enum { SUM_ALIGN = 64 };
|
||||
|
||||
bool detectSingleScale( InputArray image, Size processingRectSize,
|
||||
int yStep, double factor, std::vector<Rect>& candidates,
|
||||
std::vector<int>& rejectLevels, std::vector<double>& levelWeights,
|
||||
Size sumSize0, bool outputRejectLevels = false );
|
||||
#ifdef HAVE_OPENCL
|
||||
bool ocl_detectMultiScaleNoGrouping( const std::vector<float>& scales,
|
||||
std::vector<Rect>& candidates );
|
||||
#endif
|
||||
void detectMultiScaleNoGrouping( InputArray image, std::vector<Rect>& candidates,
|
||||
std::vector<int>& rejectLevels, std::vector<double>& levelWeights,
|
||||
double scaleFactor, Size minObjectSize, Size maxObjectSize,
|
||||
bool outputRejectLevels = false );
|
||||
|
||||
enum { MAX_FACES = 10000 };
|
||||
enum { BOOST = 0 };
|
||||
enum { DO_CANNY_PRUNING = CASCADE_DO_CANNY_PRUNING,
|
||||
SCALE_IMAGE = CASCADE_SCALE_IMAGE,
|
||||
FIND_BIGGEST_OBJECT = CASCADE_FIND_BIGGEST_OBJECT,
|
||||
DO_ROUGH_SEARCH = CASCADE_DO_ROUGH_SEARCH
|
||||
};
|
||||
|
||||
friend class CascadeClassifierInvoker;
|
||||
friend class SparseCascadeClassifierInvoker;
|
||||
|
||||
template<class FEval>
|
||||
friend int predictOrdered( CascadeClassifierImpl& cascade, Ptr<FeatureEvaluator> &featureEvaluator, double& weight);
|
||||
|
||||
template<class FEval>
|
||||
friend int predictCategorical( CascadeClassifierImpl& cascade, Ptr<FeatureEvaluator> &featureEvaluator, double& weight);
|
||||
|
||||
template<class FEval>
|
||||
friend int predictOrderedStump( CascadeClassifierImpl& cascade, Ptr<FeatureEvaluator> &featureEvaluator, double& weight);
|
||||
|
||||
template<class FEval>
|
||||
friend int predictCategoricalStump( CascadeClassifierImpl& cascade, Ptr<FeatureEvaluator> &featureEvaluator, double& weight);
|
||||
|
||||
int runAt( Ptr<FeatureEvaluator>& feval, Point pt, int scaleIdx, double& weight );
|
||||
|
||||
class Data
|
||||
{
|
||||
public:
|
||||
struct DTreeNode
|
||||
{
|
||||
int featureIdx;
|
||||
float threshold; // for ordered features only
|
||||
int left;
|
||||
int right;
|
||||
};
|
||||
|
||||
struct DTree
|
||||
{
|
||||
int nodeCount;
|
||||
};
|
||||
|
||||
struct Stage
|
||||
{
|
||||
int first;
|
||||
int ntrees;
|
||||
float threshold;
|
||||
};
|
||||
|
||||
struct Stump
|
||||
{
|
||||
Stump() : featureIdx(0), threshold(0), left(0), right(0) { }
|
||||
Stump(int _featureIdx, float _threshold, float _left, float _right)
|
||||
: featureIdx(_featureIdx), threshold(_threshold), left(_left), right(_right) {}
|
||||
|
||||
int featureIdx;
|
||||
float threshold;
|
||||
float left;
|
||||
float right;
|
||||
};
|
||||
|
||||
Data();
|
||||
|
||||
bool read(const FileNode &node);
|
||||
|
||||
int stageType;
|
||||
int featureType;
|
||||
int ncategories;
|
||||
int minNodesPerTree, maxNodesPerTree;
|
||||
Size origWinSize;
|
||||
|
||||
std::vector<Stage> stages;
|
||||
std::vector<DTree> classifiers;
|
||||
std::vector<DTreeNode> nodes;
|
||||
std::vector<float> leaves;
|
||||
std::vector<int> subsets;
|
||||
std::vector<Stump> stumps;
|
||||
};
|
||||
|
||||
Data data;
|
||||
Ptr<FeatureEvaluator> featureEvaluator;
|
||||
Ptr<CvHaarClassifierCascade> oldCascade;
|
||||
|
||||
Ptr<MaskGenerator> maskGenerator;
|
||||
UMat ugrayImage;
|
||||
UMat ufacepos, ustages, unodes, uleaves, usubsets;
|
||||
#ifdef HAVE_OPENCL
|
||||
ocl::Kernel haarKernel, lbpKernel;
|
||||
bool tryOpenCL;
|
||||
#endif
|
||||
|
||||
Mutex mtx;
|
||||
};
|
||||
|
||||
#define CC_CASCADE_PARAMS "cascadeParams"
|
||||
#define CC_STAGE_TYPE "stageType"
|
||||
#define CC_FEATURE_TYPE "featureType"
|
||||
#define CC_HEIGHT "height"
|
||||
#define CC_WIDTH "width"
|
||||
|
||||
#define CC_STAGE_NUM "stageNum"
|
||||
#define CC_STAGES "stages"
|
||||
#define CC_STAGE_PARAMS "stageParams"
|
||||
|
||||
#define CC_BOOST "BOOST"
|
||||
#define CC_MAX_DEPTH "maxDepth"
|
||||
#define CC_WEAK_COUNT "maxWeakCount"
|
||||
#define CC_STAGE_THRESHOLD "stageThreshold"
|
||||
#define CC_WEAK_CLASSIFIERS "weakClassifiers"
|
||||
#define CC_INTERNAL_NODES "internalNodes"
|
||||
#define CC_LEAF_VALUES "leafValues"
|
||||
|
||||
#define CC_FEATURES "features"
|
||||
#define CC_FEATURE_PARAMS "featureParams"
|
||||
#define CC_MAX_CAT_COUNT "maxCatCount"
|
||||
|
||||
#define CC_HAAR "HAAR"
|
||||
#define CC_RECTS "rects"
|
||||
#define CC_TILTED "tilted"
|
||||
|
||||
#define CC_LBP "LBP"
|
||||
#define CC_RECT "rect"
|
||||
|
||||
#define CC_HOG "HOG"
|
||||
|
||||
#define CV_SUM_PTRS( p0, p1, p2, p3, sum, rect, step ) \
|
||||
/* (x, y) */ \
|
||||
(p0) = sum + (rect).x + (step) * (rect).y, \
|
||||
/* (x + w, y) */ \
|
||||
(p1) = sum + (rect).x + (rect).width + (step) * (rect).y, \
|
||||
/* (x, y + h) */ \
|
||||
(p2) = sum + (rect).x + (step) * ((rect).y + (rect).height), \
|
||||
/* (x + w, y + h) */ \
|
||||
(p3) = sum + (rect).x + (rect).width + (step) * ((rect).y + (rect).height)
|
||||
|
||||
#define CV_TILTED_PTRS( p0, p1, p2, p3, tilted, rect, step ) \
|
||||
/* (x, y) */ \
|
||||
(p0) = tilted + (rect).x + (step) * (rect).y, \
|
||||
/* (x - h, y + h) */ \
|
||||
(p1) = tilted + (rect).x - (rect).height + (step) * ((rect).y + (rect).height), \
|
||||
/* (x + w, y + w) */ \
|
||||
(p2) = tilted + (rect).x + (rect).width + (step) * ((rect).y + (rect).width), \
|
||||
/* (x + w - h, y + w + h) */ \
|
||||
(p3) = tilted + (rect).x + (rect).width - (rect).height \
|
||||
+ (step) * ((rect).y + (rect).width + (rect).height)
|
||||
|
||||
#define CALC_SUM_(p0, p1, p2, p3, offset) \
|
||||
((p0)[offset] - (p1)[offset] - (p2)[offset] + (p3)[offset])
|
||||
|
||||
#define CALC_SUM(rect,offset) CALC_SUM_((rect)[0], (rect)[1], (rect)[2], (rect)[3], offset)
|
||||
|
||||
#define CV_SUM_OFS( p0, p1, p2, p3, sum, rect, step ) \
|
||||
/* (x, y) */ \
|
||||
(p0) = sum + (rect).x + (step) * (rect).y, \
|
||||
/* (x + w, y) */ \
|
||||
(p1) = sum + (rect).x + (rect).width + (step) * (rect).y, \
|
||||
/* (x, y + h) */ \
|
||||
(p2) = sum + (rect).x + (step) * ((rect).y + (rect).height), \
|
||||
/* (x + w, y + h) */ \
|
||||
(p3) = sum + (rect).x + (rect).width + (step) * ((rect).y + (rect).height)
|
||||
|
||||
#define CV_TILTED_OFS( p0, p1, p2, p3, tilted, rect, step ) \
|
||||
/* (x, y) */ \
|
||||
(p0) = tilted + (rect).x + (step) * (rect).y, \
|
||||
/* (x - h, y + h) */ \
|
||||
(p1) = tilted + (rect).x - (rect).height + (step) * ((rect).y + (rect).height), \
|
||||
/* (x + w, y + w) */ \
|
||||
(p2) = tilted + (rect).x + (rect).width + (step) * ((rect).y + (rect).width), \
|
||||
/* (x + w - h, y + w + h) */ \
|
||||
(p3) = tilted + (rect).x + (rect).width - (rect).height \
|
||||
+ (step) * ((rect).y + (rect).width + (rect).height)
|
||||
|
||||
#define CALC_SUM_OFS_(p0, p1, p2, p3, ptr) \
|
||||
((ptr)[p0] - (ptr)[p1] - (ptr)[p2] + (ptr)[p3])
|
||||
|
||||
#define CALC_SUM_OFS(rect, ptr) CALC_SUM_OFS_((rect)[0], (rect)[1], (rect)[2], (rect)[3], ptr)
|
||||
|
||||
//---------------------------------------------- HaarEvaluator ---------------------------------------
|
||||
class HaarEvaluator CV_FINAL : public FeatureEvaluator
|
||||
{
|
||||
public:
|
||||
struct Feature
|
||||
{
|
||||
Feature();
|
||||
bool read(const FileNode& node, const Size& origWinSize);
|
||||
|
||||
bool tilted;
|
||||
|
||||
enum { RECT_NUM = 3 };
|
||||
struct RectWeigth
|
||||
{
|
||||
Rect r;
|
||||
float weight;
|
||||
} rect[RECT_NUM];
|
||||
};
|
||||
|
||||
struct OptFeature
|
||||
{
|
||||
OptFeature();
|
||||
|
||||
enum { RECT_NUM = Feature::RECT_NUM };
|
||||
float calc( const int* pwin ) const;
|
||||
void setOffsets( const Feature& _f, int step, int tofs );
|
||||
|
||||
int ofs[RECT_NUM][4];
|
||||
float weight[4];
|
||||
};
|
||||
|
||||
HaarEvaluator();
|
||||
virtual ~HaarEvaluator() CV_OVERRIDE;
|
||||
|
||||
virtual bool read( const FileNode& node, Size origWinSize) CV_OVERRIDE;
|
||||
virtual Ptr<FeatureEvaluator> clone() const CV_OVERRIDE;
|
||||
virtual int getFeatureType() const CV_OVERRIDE { return FeatureEvaluator::HAAR; }
|
||||
|
||||
virtual bool setWindow(Point p, int scaleIdx) CV_OVERRIDE;
|
||||
Rect getNormRect() const;
|
||||
int getSquaresOffset() const;
|
||||
|
||||
float operator()(int featureIdx) const
|
||||
{ return optfeaturesPtr[featureIdx].calc(pwin) * varianceNormFactor; }
|
||||
virtual float calcOrd(int featureIdx) const CV_OVERRIDE
|
||||
{ return (*this)(featureIdx); }
|
||||
|
||||
protected:
|
||||
virtual void computeChannels( int i, InputArray img ) CV_OVERRIDE;
|
||||
virtual void computeOptFeatures() CV_OVERRIDE;
|
||||
|
||||
Ptr<std::vector<Feature> > features;
|
||||
Ptr<std::vector<OptFeature> > optfeatures;
|
||||
Ptr<std::vector<OptFeature> > optfeatures_lbuf;
|
||||
bool hasTiltedFeatures;
|
||||
|
||||
int tofs, sqofs;
|
||||
Vec4i nofs;
|
||||
Rect normrect;
|
||||
const int* pwin;
|
||||
OptFeature* optfeaturesPtr; // optimization
|
||||
float varianceNormFactor;
|
||||
};
|
||||
|
||||
inline HaarEvaluator::Feature :: Feature()
|
||||
{
|
||||
tilted = false;
|
||||
rect[0].r = rect[1].r = rect[2].r = Rect();
|
||||
rect[0].weight = rect[1].weight = rect[2].weight = 0;
|
||||
}
|
||||
|
||||
inline HaarEvaluator::OptFeature :: OptFeature()
|
||||
{
|
||||
weight[0] = weight[1] = weight[2] = 0.f;
|
||||
|
||||
ofs[0][0] = ofs[0][1] = ofs[0][2] = ofs[0][3] =
|
||||
ofs[1][0] = ofs[1][1] = ofs[1][2] = ofs[1][3] =
|
||||
ofs[2][0] = ofs[2][1] = ofs[2][2] = ofs[2][3] = 0;
|
||||
}
|
||||
|
||||
inline float HaarEvaluator::OptFeature :: calc( const int* ptr ) const
|
||||
{
|
||||
float ret = weight[0] * CALC_SUM_OFS(ofs[0], ptr) +
|
||||
weight[1] * CALC_SUM_OFS(ofs[1], ptr);
|
||||
|
||||
if( weight[2] != 0.0f )
|
||||
ret += weight[2] * CALC_SUM_OFS(ofs[2], ptr);
|
||||
|
||||
return ret;
|
||||
}
|
||||
|
||||
//---------------------------------------------- LBPEvaluator -------------------------------------
|
||||
|
||||
class LBPEvaluator CV_FINAL : public FeatureEvaluator
|
||||
{
|
||||
public:
|
||||
struct Feature
|
||||
{
|
||||
Feature();
|
||||
Feature( int x, int y, int _block_w, int _block_h ) :
|
||||
rect(x, y, _block_w, _block_h) {}
|
||||
|
||||
bool read(const FileNode& node, const Size& origWinSize);
|
||||
|
||||
Rect rect; // weight and height for block
|
||||
};
|
||||
|
||||
struct OptFeature
|
||||
{
|
||||
OptFeature();
|
||||
|
||||
int calc( const int* pwin ) const;
|
||||
void setOffsets( const Feature& _f, int step );
|
||||
int ofs[16];
|
||||
};
|
||||
|
||||
LBPEvaluator();
|
||||
virtual ~LBPEvaluator() CV_OVERRIDE;
|
||||
|
||||
virtual bool read( const FileNode& node, Size origWinSize ) CV_OVERRIDE;
|
||||
virtual Ptr<FeatureEvaluator> clone() const CV_OVERRIDE;
|
||||
virtual int getFeatureType() const CV_OVERRIDE { return FeatureEvaluator::LBP; }
|
||||
|
||||
virtual bool setWindow(Point p, int scaleIdx) CV_OVERRIDE;
|
||||
|
||||
int operator()(int featureIdx) const
|
||||
{ return optfeaturesPtr[featureIdx].calc(pwin); }
|
||||
virtual int calcCat(int featureIdx) const CV_OVERRIDE
|
||||
{ return (*this)(featureIdx); }
|
||||
protected:
|
||||
virtual void computeChannels( int i, InputArray img ) CV_OVERRIDE;
|
||||
virtual void computeOptFeatures() CV_OVERRIDE;
|
||||
|
||||
Ptr<std::vector<Feature> > features;
|
||||
Ptr<std::vector<OptFeature> > optfeatures;
|
||||
Ptr<std::vector<OptFeature> > optfeatures_lbuf;
|
||||
OptFeature* optfeaturesPtr; // optimization
|
||||
|
||||
const int* pwin;
|
||||
};
|
||||
|
||||
|
||||
inline LBPEvaluator::Feature :: Feature()
|
||||
{
|
||||
rect = Rect();
|
||||
}
|
||||
|
||||
inline LBPEvaluator::OptFeature :: OptFeature()
|
||||
{
|
||||
for( int i = 0; i < 16; i++ )
|
||||
ofs[i] = 0;
|
||||
}
|
||||
|
||||
inline int LBPEvaluator::OptFeature :: calc( const int* p ) const
|
||||
{
|
||||
int cval = CALC_SUM_OFS_( ofs[5], ofs[6], ofs[9], ofs[10], p );
|
||||
|
||||
return (CALC_SUM_OFS_( ofs[0], ofs[1], ofs[4], ofs[5], p ) >= cval ? 128 : 0) | // 0
|
||||
(CALC_SUM_OFS_( ofs[1], ofs[2], ofs[5], ofs[6], p ) >= cval ? 64 : 0) | // 1
|
||||
(CALC_SUM_OFS_( ofs[2], ofs[3], ofs[6], ofs[7], p ) >= cval ? 32 : 0) | // 2
|
||||
(CALC_SUM_OFS_( ofs[6], ofs[7], ofs[10], ofs[11], p ) >= cval ? 16 : 0) | // 5
|
||||
(CALC_SUM_OFS_( ofs[10], ofs[11], ofs[14], ofs[15], p ) >= cval ? 8 : 0)| // 8
|
||||
(CALC_SUM_OFS_( ofs[9], ofs[10], ofs[13], ofs[14], p ) >= cval ? 4 : 0)| // 7
|
||||
(CALC_SUM_OFS_( ofs[8], ofs[9], ofs[12], ofs[13], p ) >= cval ? 2 : 0)| // 6
|
||||
(CALC_SUM_OFS_( ofs[4], ofs[5], ofs[8], ofs[9], p ) >= cval ? 1 : 0);
|
||||
}
|
||||
|
||||
|
||||
//---------------------------------------------- predictor functions -------------------------------------
|
||||
|
||||
template<class FEval>
|
||||
inline int predictOrdered( CascadeClassifierImpl& cascade,
|
||||
Ptr<FeatureEvaluator> &_featureEvaluator, double& sum )
|
||||
{
|
||||
CV_INSTRUMENT_REGION();
|
||||
|
||||
int nstages = (int)cascade.data.stages.size();
|
||||
int nodeOfs = 0, leafOfs = 0;
|
||||
FEval& featureEvaluator = (FEval&)*_featureEvaluator;
|
||||
float* cascadeLeaves = &cascade.data.leaves[0];
|
||||
CascadeClassifierImpl::Data::DTreeNode* cascadeNodes = &cascade.data.nodes[0];
|
||||
CascadeClassifierImpl::Data::DTree* cascadeWeaks = &cascade.data.classifiers[0];
|
||||
CascadeClassifierImpl::Data::Stage* cascadeStages = &cascade.data.stages[0];
|
||||
|
||||
for( int si = 0; si < nstages; si++ )
|
||||
{
|
||||
CascadeClassifierImpl::Data::Stage& stage = cascadeStages[si];
|
||||
int wi, ntrees = stage.ntrees;
|
||||
sum = 0;
|
||||
|
||||
for( wi = 0; wi < ntrees; wi++ )
|
||||
{
|
||||
CascadeClassifierImpl::Data::DTree& weak = cascadeWeaks[stage.first + wi];
|
||||
int idx = 0, root = nodeOfs;
|
||||
|
||||
do
|
||||
{
|
||||
CascadeClassifierImpl::Data::DTreeNode& node = cascadeNodes[root + idx];
|
||||
double val = featureEvaluator(node.featureIdx);
|
||||
idx = val < node.threshold ? node.left : node.right;
|
||||
}
|
||||
while( idx > 0 );
|
||||
sum += cascadeLeaves[leafOfs - idx];
|
||||
nodeOfs += weak.nodeCount;
|
||||
leafOfs += weak.nodeCount + 1;
|
||||
}
|
||||
if( sum < stage.threshold )
|
||||
return -si;
|
||||
}
|
||||
return 1;
|
||||
}
|
||||
|
||||
template<class FEval>
|
||||
inline int predictCategorical( CascadeClassifierImpl& cascade,
|
||||
Ptr<FeatureEvaluator> &_featureEvaluator, double& sum )
|
||||
{
|
||||
CV_INSTRUMENT_REGION();
|
||||
|
||||
int nstages = (int)cascade.data.stages.size();
|
||||
int nodeOfs = 0, leafOfs = 0;
|
||||
FEval& featureEvaluator = (FEval&)*_featureEvaluator;
|
||||
size_t subsetSize = (cascade.data.ncategories + 31)/32;
|
||||
int* cascadeSubsets = &cascade.data.subsets[0];
|
||||
float* cascadeLeaves = &cascade.data.leaves[0];
|
||||
CascadeClassifierImpl::Data::DTreeNode* cascadeNodes = &cascade.data.nodes[0];
|
||||
CascadeClassifierImpl::Data::DTree* cascadeWeaks = &cascade.data.classifiers[0];
|
||||
CascadeClassifierImpl::Data::Stage* cascadeStages = &cascade.data.stages[0];
|
||||
|
||||
for(int si = 0; si < nstages; si++ )
|
||||
{
|
||||
CascadeClassifierImpl::Data::Stage& stage = cascadeStages[si];
|
||||
int wi, ntrees = stage.ntrees;
|
||||
sum = 0;
|
||||
|
||||
for( wi = 0; wi < ntrees; wi++ )
|
||||
{
|
||||
CascadeClassifierImpl::Data::DTree& weak = cascadeWeaks[stage.first + wi];
|
||||
int idx = 0, root = nodeOfs;
|
||||
do
|
||||
{
|
||||
CascadeClassifierImpl::Data::DTreeNode& node = cascadeNodes[root + idx];
|
||||
int c = featureEvaluator(node.featureIdx);
|
||||
const int* subset = &cascadeSubsets[(root + idx)*subsetSize];
|
||||
idx = (subset[c>>5] & (1 << (c & 31))) ? node.left : node.right;
|
||||
}
|
||||
while( idx > 0 );
|
||||
sum += cascadeLeaves[leafOfs - idx];
|
||||
nodeOfs += weak.nodeCount;
|
||||
leafOfs += weak.nodeCount + 1;
|
||||
}
|
||||
if( sum < stage.threshold )
|
||||
return -si;
|
||||
}
|
||||
return 1;
|
||||
}
|
||||
|
||||
template<class FEval>
|
||||
inline int predictOrderedStump( CascadeClassifierImpl& cascade,
|
||||
Ptr<FeatureEvaluator> &_featureEvaluator, double& sum )
|
||||
{
|
||||
CV_INSTRUMENT_REGION();
|
||||
|
||||
CV_Assert(!cascade.data.stumps.empty());
|
||||
FEval& featureEvaluator = (FEval&)*_featureEvaluator;
|
||||
const CascadeClassifierImpl::Data::Stump* cascadeStumps = &cascade.data.stumps[0];
|
||||
const CascadeClassifierImpl::Data::Stage* cascadeStages = &cascade.data.stages[0];
|
||||
|
||||
int nstages = (int)cascade.data.stages.size();
|
||||
double tmp = 0;
|
||||
|
||||
for( int stageIdx = 0; stageIdx < nstages; stageIdx++ )
|
||||
{
|
||||
const CascadeClassifierImpl::Data::Stage& stage = cascadeStages[stageIdx];
|
||||
tmp = 0;
|
||||
|
||||
int ntrees = stage.ntrees;
|
||||
for( int i = 0; i < ntrees; i++ )
|
||||
{
|
||||
const CascadeClassifierImpl::Data::Stump& stump = cascadeStumps[i];
|
||||
double value = featureEvaluator(stump.featureIdx);
|
||||
tmp += value < stump.threshold ? stump.left : stump.right;
|
||||
}
|
||||
|
||||
if( tmp < stage.threshold )
|
||||
{
|
||||
sum = (double)tmp;
|
||||
return -stageIdx;
|
||||
}
|
||||
cascadeStumps += ntrees;
|
||||
}
|
||||
|
||||
sum = (double)tmp;
|
||||
return 1;
|
||||
}
|
||||
|
||||
template<class FEval>
|
||||
inline int predictCategoricalStump( CascadeClassifierImpl& cascade,
|
||||
Ptr<FeatureEvaluator> &_featureEvaluator, double& sum )
|
||||
{
|
||||
CV_INSTRUMENT_REGION();
|
||||
|
||||
CV_Assert(!cascade.data.stumps.empty());
|
||||
int nstages = (int)cascade.data.stages.size();
|
||||
FEval& featureEvaluator = (FEval&)*_featureEvaluator;
|
||||
size_t subsetSize = (cascade.data.ncategories + 31)/32;
|
||||
const int* cascadeSubsets = &cascade.data.subsets[0];
|
||||
const CascadeClassifierImpl::Data::Stump* cascadeStumps = &cascade.data.stumps[0];
|
||||
const CascadeClassifierImpl::Data::Stage* cascadeStages = &cascade.data.stages[0];
|
||||
|
||||
double tmp = 0;
|
||||
for( int si = 0; si < nstages; si++ )
|
||||
{
|
||||
const CascadeClassifierImpl::Data::Stage& stage = cascadeStages[si];
|
||||
int wi, ntrees = stage.ntrees;
|
||||
tmp = 0;
|
||||
|
||||
for( wi = 0; wi < ntrees; wi++ )
|
||||
{
|
||||
const CascadeClassifierImpl::Data::Stump& stump = cascadeStumps[wi];
|
||||
int c = featureEvaluator(stump.featureIdx);
|
||||
const int* subset = &cascadeSubsets[wi*subsetSize];
|
||||
tmp += (subset[c>>5] & (1 << (c & 31))) ? stump.left : stump.right;
|
||||
}
|
||||
|
||||
if( tmp < stage.threshold )
|
||||
{
|
||||
sum = tmp;
|
||||
return -si;
|
||||
}
|
||||
|
||||
cascadeStumps += ntrees;
|
||||
cascadeSubsets += ntrees*subsetSize;
|
||||
}
|
||||
|
||||
sum = (double)tmp;
|
||||
return 1;
|
||||
}
|
||||
|
||||
namespace haar_cvt
|
||||
{
|
||||
bool convert(const FileNode& oldcascade_root, FileStorage& newfs);
|
||||
}
|
||||
|
||||
}
|
||||
@@ -0,0 +1,273 @@
|
||||
/*M///////////////////////////////////////////////////////////////////////////////////////
|
||||
//
|
||||
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
|
||||
//
|
||||
// By downloading, copying, installing or using the software you agree to this license.
|
||||
// If you do not agree to this license, do not download, install,
|
||||
// copy or use the software.
|
||||
//
|
||||
//
|
||||
// License Agreement
|
||||
// For Open Source Computer Vision Library
|
||||
//
|
||||
// Copyright (C) 2013, Itseez Inc, all rights reserved.
|
||||
// Third party copyrights are property of their respective owners.
|
||||
//
|
||||
// Redistribution and use in source and binary forms, with or without modification,
|
||||
// are permitted provided that the following conditions are met:
|
||||
//
|
||||
// * Redistribution's of source code must retain the above copyright notice,
|
||||
// this list of conditions and the following disclaimer.
|
||||
//
|
||||
// * Redistribution's in binary form must reproduce the above copyright notice,
|
||||
// this list of conditions and the following disclaimer in the documentation
|
||||
// and/or other materials provided with the distribution.
|
||||
//
|
||||
// * The name of Intel Corporation may not be used to endorse or promote products
|
||||
// derived from this software without specific prior written permission.
|
||||
//
|
||||
// This software is provided by the copyright holders and contributors "as is" and
|
||||
// any express or implied warranties, including, but not limited to, the implied
|
||||
// warranties of merchantability and fitness for a particular purpose are disclaimed.
|
||||
// In no event shall the Intel Corporation or contributors be liable for any direct,
|
||||
// indirect, incidental, special, exemplary, or consequential damages
|
||||
// (including, but not limited to, procurement of substitute goods or services;
|
||||
// loss of use, data, or profits; or business interruption) however caused
|
||||
// and on any theory of liability, whether in contract, strict liability,
|
||||
// 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.
|
||||
//
|
||||
//M*/
|
||||
|
||||
/* Haar features calculation */
|
||||
|
||||
#include "precomp.hpp"
|
||||
#include "cascadedetect.hpp"
|
||||
#include <stdio.h>
|
||||
|
||||
namespace cv
|
||||
{
|
||||
|
||||
/* field names */
|
||||
|
||||
#define ICV_HAAR_SIZE_NAME "size"
|
||||
#define ICV_HAAR_STAGES_NAME "stages"
|
||||
#define ICV_HAAR_TREES_NAME "trees"
|
||||
#define ICV_HAAR_FEATURE_NAME "feature"
|
||||
#define ICV_HAAR_RECTS_NAME "rects"
|
||||
#define ICV_HAAR_TILTED_NAME "tilted"
|
||||
#define ICV_HAAR_THRESHOLD_NAME "threshold"
|
||||
#define ICV_HAAR_LEFT_NODE_NAME "left_node"
|
||||
#define ICV_HAAR_LEFT_VAL_NAME "left_val"
|
||||
#define ICV_HAAR_RIGHT_NODE_NAME "right_node"
|
||||
#define ICV_HAAR_RIGHT_VAL_NAME "right_val"
|
||||
#define ICV_HAAR_STAGE_THRESHOLD_NAME "stage_threshold"
|
||||
#define ICV_HAAR_PARENT_NAME "parent"
|
||||
#define ICV_HAAR_NEXT_NAME "next"
|
||||
|
||||
namespace haar_cvt
|
||||
{
|
||||
|
||||
struct HaarFeature
|
||||
{
|
||||
enum { RECT_NUM = 3 };
|
||||
|
||||
HaarFeature()
|
||||
{
|
||||
tilted = false;
|
||||
for( int i = 0; i < RECT_NUM; i++ )
|
||||
{
|
||||
rect[i].r = Rect(0,0,0,0);
|
||||
rect[i].weight = 0.f;
|
||||
}
|
||||
}
|
||||
bool tilted;
|
||||
struct
|
||||
{
|
||||
Rect r;
|
||||
float weight;
|
||||
} rect[RECT_NUM];
|
||||
};
|
||||
|
||||
struct HaarClassifierNode
|
||||
{
|
||||
HaarClassifierNode()
|
||||
{
|
||||
f = left = right = 0;
|
||||
threshold = 0.f;
|
||||
}
|
||||
int f, left, right;
|
||||
float threshold;
|
||||
};
|
||||
|
||||
struct HaarClassifier
|
||||
{
|
||||
std::vector<HaarClassifierNode> nodes;
|
||||
std::vector<float> leaves;
|
||||
};
|
||||
|
||||
struct HaarStageClassifier
|
||||
{
|
||||
double threshold;
|
||||
std::vector<HaarClassifier> weaks;
|
||||
};
|
||||
|
||||
bool convert(const FileNode& oldroot, FileStorage& newfs)
|
||||
{
|
||||
FileNode sznode = oldroot[ICV_HAAR_SIZE_NAME];
|
||||
if( sznode.empty() )
|
||||
return false;
|
||||
Size cascadesize;
|
||||
cascadesize.width = (int)sznode[0];
|
||||
cascadesize.height = (int)sznode[1];
|
||||
std::vector<HaarFeature> features;
|
||||
|
||||
int i, j, k, n;
|
||||
|
||||
FileNode stages_seq = oldroot[ICV_HAAR_STAGES_NAME];
|
||||
int nstages = (int)stages_seq.size();
|
||||
std::vector<HaarStageClassifier> stages(nstages);
|
||||
|
||||
for( i = 0; i < nstages; i++ )
|
||||
{
|
||||
FileNode stagenode = stages_seq[i];
|
||||
HaarStageClassifier& stage = stages[i];
|
||||
stage.threshold = (double)stagenode[ICV_HAAR_STAGE_THRESHOLD_NAME];
|
||||
FileNode weaks_seq = stagenode[ICV_HAAR_TREES_NAME];
|
||||
int nweaks = (int)weaks_seq.size();
|
||||
stage.weaks.resize(nweaks);
|
||||
|
||||
for( j = 0; j < nweaks; j++ )
|
||||
{
|
||||
HaarClassifier& weak = stage.weaks[j];
|
||||
FileNode weaknode = weaks_seq[j];
|
||||
int nnodes = (int)weaknode.size();
|
||||
|
||||
for( n = 0; n < nnodes; n++ )
|
||||
{
|
||||
FileNode nnode = weaknode[n];
|
||||
FileNode fnode = nnode[ICV_HAAR_FEATURE_NAME];
|
||||
HaarFeature f;
|
||||
HaarClassifierNode node;
|
||||
node.f = (int)features.size();
|
||||
f.tilted = (int)fnode[ICV_HAAR_TILTED_NAME] != 0;
|
||||
FileNode rects_seq = fnode[ICV_HAAR_RECTS_NAME];
|
||||
int nrects = (int)rects_seq.size();
|
||||
|
||||
for( k = 0; k < nrects; k++ )
|
||||
{
|
||||
FileNode rnode = rects_seq[k];
|
||||
f.rect[k].r.x = (int)rnode[0];
|
||||
f.rect[k].r.y = (int)rnode[1];
|
||||
f.rect[k].r.width = (int)rnode[2];
|
||||
f.rect[k].r.height = (int)rnode[3];
|
||||
f.rect[k].weight = (float)rnode[4];
|
||||
}
|
||||
features.push_back(f);
|
||||
node.threshold = nnode[ICV_HAAR_THRESHOLD_NAME];
|
||||
FileNode leftValNode = nnode[ICV_HAAR_LEFT_VAL_NAME];
|
||||
if( !leftValNode.empty() )
|
||||
{
|
||||
node.left = -(int)weak.leaves.size();
|
||||
weak.leaves.push_back((float)leftValNode);
|
||||
}
|
||||
else
|
||||
{
|
||||
node.left = (int)nnode[ICV_HAAR_LEFT_NODE_NAME];
|
||||
}
|
||||
FileNode rightValNode = nnode[ICV_HAAR_RIGHT_VAL_NAME];
|
||||
if( !rightValNode.empty() )
|
||||
{
|
||||
node.right = -(int)weak.leaves.size();
|
||||
weak.leaves.push_back((float)rightValNode);
|
||||
}
|
||||
else
|
||||
{
|
||||
node.right = (int)nnode[ICV_HAAR_RIGHT_NODE_NAME];
|
||||
}
|
||||
weak.nodes.push_back(node);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
int maxWeakCount = 0, nfeatures = (int)features.size();
|
||||
for( i = 0; i < nstages; i++ )
|
||||
maxWeakCount = std::max(maxWeakCount, (int)stages[i].weaks.size());
|
||||
|
||||
newfs << "cascade" << "{:opencv-cascade-classifier"
|
||||
<< "stageType" << "BOOST"
|
||||
<< "featureType" << "HAAR"
|
||||
<< "height" << cascadesize.width
|
||||
<< "width" << cascadesize.height
|
||||
<< "stageParams" << "{"
|
||||
<< "maxWeakCount" << (int)maxWeakCount
|
||||
<< "}"
|
||||
<< "featureParams" << "{"
|
||||
<< "maxCatCount" << 0
|
||||
<< "}"
|
||||
<< "stageNum" << (int)nstages
|
||||
<< "stages" << "[";
|
||||
|
||||
for( i = 0; i < nstages; i++ )
|
||||
{
|
||||
int nweaks = (int)stages[i].weaks.size();
|
||||
newfs << "{" << "maxWeakCount" << (int)nweaks
|
||||
<< "stageThreshold" << stages[i].threshold
|
||||
<< "weakClassifiers" << "[";
|
||||
for( j = 0; j < nweaks; j++ )
|
||||
{
|
||||
const HaarClassifier& c = stages[i].weaks[j];
|
||||
newfs << "{" << "internalNodes" << "[:";
|
||||
int nnodes = (int)c.nodes.size(), nleaves = (int)c.leaves.size();
|
||||
for( k = 0; k < nnodes; k++ )
|
||||
newfs << c.nodes[k].left << c.nodes[k].right
|
||||
<< c.nodes[k].f << c.nodes[k].threshold;
|
||||
newfs << "]" << "leafValues" << "[:";
|
||||
for( k = 0; k < nleaves; k++ )
|
||||
newfs << c.leaves[k];
|
||||
newfs << "]" << "}";
|
||||
}
|
||||
newfs << "]" << "}";
|
||||
}
|
||||
|
||||
newfs << "]"
|
||||
<< "features" << "[";
|
||||
|
||||
for( i = 0; i < nfeatures; i++ )
|
||||
{
|
||||
const HaarFeature& f = features[i];
|
||||
newfs << "{" << "rects" << "[";
|
||||
for( j = 0; j < HaarFeature::RECT_NUM; j++ )
|
||||
{
|
||||
if( j >= 2 && fabs(f.rect[j].weight) < FLT_EPSILON )
|
||||
break;
|
||||
newfs << "[:" << f.rect[j].r.x << f.rect[j].r.y <<
|
||||
f.rect[j].r.width << f.rect[j].r.height << f.rect[j].weight << "]";
|
||||
}
|
||||
newfs << "]";
|
||||
if( f.tilted )
|
||||
newfs << "tilted" << 1;
|
||||
newfs << "}";
|
||||
}
|
||||
|
||||
newfs << "]" << "}";
|
||||
return true;
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
bool CascadeClassifier::convert(const String& oldcascade, const String& newcascade)
|
||||
{
|
||||
FileStorage oldfs(oldcascade, FileStorage::READ);
|
||||
FileStorage newfs(newcascade, FileStorage::WRITE);
|
||||
if( !oldfs.isOpened() || !newfs.isOpened() )
|
||||
return false;
|
||||
FileNode oldroot = oldfs.getFirstTopLevelNode();
|
||||
|
||||
bool ok = haar_cvt::convert(oldroot, newfs);
|
||||
if( !ok && newcascade.size() > 0 )
|
||||
remove(newcascade.c_str());
|
||||
return ok;
|
||||
}
|
||||
|
||||
}
|
||||
@@ -0,0 +1,885 @@
|
||||
/*M///////////////////////////////////////////////////////////////////////////////////////
|
||||
//
|
||||
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
|
||||
//
|
||||
// By downloading, copying, installing or using the software you agree to this license.
|
||||
// If you do not agree to this license, do not download, install,
|
||||
// copy or use the software.
|
||||
//
|
||||
//
|
||||
// License Agreement
|
||||
// For Open Source Computer Vision Library
|
||||
//
|
||||
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
|
||||
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
|
||||
// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
|
||||
// Third party copyrights are property of their respective owners.
|
||||
//
|
||||
// Redistribution and use in source and binary forms, with or without modification,
|
||||
// are permitted provided that the following conditions are met:
|
||||
//
|
||||
// * Redistribution's of source code must retain the above copyright notice,
|
||||
// this list of conditions and the following disclaimer.
|
||||
//
|
||||
// * Redistribution's in binary form must reproduce the above copyright notice,
|
||||
// this list of conditions and the following disclaimer in the documentation
|
||||
// and/or other materials provided with the distribution.
|
||||
//
|
||||
// * The name of the copyright holders may not be used to endorse or promote products
|
||||
// derived from this software without specific prior written permission.
|
||||
//
|
||||
// This software is provided by the copyright holders and contributors "as is" and
|
||||
// any express or implied warranties, including, but not limited to, the implied
|
||||
// warranties of merchantability and fitness for a particular purpose are disclaimed.
|
||||
// In no event shall the Intel Corporation or contributors be liable for any direct,
|
||||
// indirect, incidental, special, exemplary, or consequential damages
|
||||
// (including, but not limited to, procurement of substitute goods or services;
|
||||
// loss of use, data, or profits; or business interruption) however caused
|
||||
// and on any theory of liability, whether in contract, strict liability,
|
||||
// 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.
|
||||
//
|
||||
//M*/
|
||||
|
||||
#include "precomp.hpp"
|
||||
|
||||
#include "opencv2/core/utility.hpp"
|
||||
|
||||
#include <thread>
|
||||
#include <mutex>
|
||||
#include <condition_variable>
|
||||
|
||||
#if defined(DEBUG) || defined(_DEBUG)
|
||||
#undef DEBUGLOGS
|
||||
#define DEBUGLOGS 1
|
||||
#endif
|
||||
|
||||
#ifndef DEBUGLOGS
|
||||
#define DEBUGLOGS 0
|
||||
#endif
|
||||
|
||||
#ifdef __ANDROID__
|
||||
#include <android/log.h>
|
||||
#define LOG_TAG "OBJECT_DETECTOR"
|
||||
#define LOGD0(...) ((void)__android_log_print(ANDROID_LOG_DEBUG, LOG_TAG, __VA_ARGS__))
|
||||
#define LOGI0(...) ((void)__android_log_print(ANDROID_LOG_INFO, LOG_TAG, __VA_ARGS__))
|
||||
#define LOGW0(...) ((void)__android_log_print(ANDROID_LOG_WARN, LOG_TAG, __VA_ARGS__))
|
||||
#define LOGE0(...) ((void)__android_log_print(ANDROID_LOG_ERROR, LOG_TAG, __VA_ARGS__))
|
||||
#else
|
||||
|
||||
#include <stdio.h>
|
||||
|
||||
#define LOGD0(_str, ...) (printf(_str , ## __VA_ARGS__), printf("\n"), fflush(stdout))
|
||||
#define LOGI0(_str, ...) (printf(_str , ## __VA_ARGS__), printf("\n"), fflush(stdout))
|
||||
#define LOGW0(_str, ...) (printf(_str , ## __VA_ARGS__), printf("\n"), fflush(stdout))
|
||||
#define LOGE0(_str, ...) (printf(_str , ## __VA_ARGS__), printf("\n"), fflush(stdout))
|
||||
#endif //__ANDROID__
|
||||
|
||||
#if DEBUGLOGS
|
||||
#define LOGD(_str, ...) LOGD0(_str , ## __VA_ARGS__)
|
||||
#define LOGI(_str, ...) LOGI0(_str , ## __VA_ARGS__)
|
||||
#define LOGW(_str, ...) LOGW0(_str , ## __VA_ARGS__)
|
||||
#define LOGE(_str, ...) LOGE0(_str , ## __VA_ARGS__)
|
||||
#else
|
||||
#define LOGD(...)
|
||||
#define LOGI(...)
|
||||
#define LOGW(...)
|
||||
#define LOGE(...)
|
||||
#endif //DEBUGLOGS
|
||||
|
||||
|
||||
using namespace cv;
|
||||
|
||||
static inline cv::Point2f centerRect(const cv::Rect& r)
|
||||
{
|
||||
return cv::Point2f(r.x+((float)r.width)/2, r.y+((float)r.height)/2);
|
||||
}
|
||||
|
||||
static inline cv::Rect scale_rect(const cv::Rect& r, float scale)
|
||||
{
|
||||
cv::Point2f m=centerRect(r);
|
||||
float width = r.width * scale;
|
||||
float height = r.height * scale;
|
||||
int x=cvRound(m.x - width/2);
|
||||
int y=cvRound(m.y - height/2);
|
||||
|
||||
return cv::Rect(x, y, cvRound(width), cvRound(height));
|
||||
}
|
||||
|
||||
namespace cv
|
||||
{
|
||||
void* workcycleObjectDetectorFunction(void* p);
|
||||
}
|
||||
|
||||
class cv::DetectionBasedTracker::SeparateDetectionWork
|
||||
{
|
||||
public:
|
||||
SeparateDetectionWork(cv::DetectionBasedTracker& _detectionBasedTracker, cv::Ptr<DetectionBasedTracker::IDetector> _detector,
|
||||
const cv::DetectionBasedTracker::Parameters& params);
|
||||
virtual ~SeparateDetectionWork();
|
||||
bool communicateWithDetectingThread(const Mat& imageGray, std::vector<Rect>& rectsWhereRegions);
|
||||
bool run();
|
||||
void stop();
|
||||
void resetTracking();
|
||||
|
||||
inline bool isWorking()
|
||||
{
|
||||
return (stateThread==STATE_THREAD_WORKING_SLEEPING) || (stateThread==STATE_THREAD_WORKING_WITH_IMAGE);
|
||||
}
|
||||
void setParameters(const cv::DetectionBasedTracker::Parameters& params)
|
||||
{
|
||||
std::unique_lock<std::mutex> mtx_lock(mtx);
|
||||
parameters = params;
|
||||
}
|
||||
|
||||
inline void init()
|
||||
{
|
||||
std::unique_lock<std::mutex> mtx_lock(mtx);
|
||||
stateThread = STATE_THREAD_STOPPED;
|
||||
isObjectDetectingReady = false;
|
||||
shouldObjectDetectingResultsBeForgot = false;
|
||||
objectDetectorThreadStartStop.notify_one();
|
||||
}
|
||||
protected:
|
||||
|
||||
DetectionBasedTracker& detectionBasedTracker;
|
||||
cv::Ptr<DetectionBasedTracker::IDetector> cascadeInThread;
|
||||
std::thread second_workthread;
|
||||
std::mutex mtx;
|
||||
std::condition_variable objectDetectorRun;
|
||||
std::condition_variable objectDetectorThreadStartStop;
|
||||
std::vector<cv::Rect> resultDetect;
|
||||
volatile bool isObjectDetectingReady;
|
||||
volatile bool shouldObjectDetectingResultsBeForgot;
|
||||
|
||||
enum StateSeparatedThread {
|
||||
STATE_THREAD_STOPPED=0,
|
||||
STATE_THREAD_WORKING_SLEEPING,
|
||||
STATE_THREAD_WORKING_WITH_IMAGE,
|
||||
STATE_THREAD_WORKING,
|
||||
STATE_THREAD_STOPPING
|
||||
};
|
||||
volatile StateSeparatedThread stateThread;
|
||||
|
||||
cv::Mat imageSeparateDetecting;
|
||||
|
||||
void workcycleObjectDetector();
|
||||
friend void* workcycleObjectDetectorFunction(void* p);
|
||||
|
||||
long long timeWhenDetectingThreadStartedWork;
|
||||
cv::DetectionBasedTracker::Parameters parameters;
|
||||
};
|
||||
|
||||
cv::DetectionBasedTracker::SeparateDetectionWork::SeparateDetectionWork(DetectionBasedTracker& _detectionBasedTracker, cv::Ptr<DetectionBasedTracker::IDetector> _detector,
|
||||
const cv::DetectionBasedTracker::Parameters& params)
|
||||
:detectionBasedTracker(_detectionBasedTracker),
|
||||
cascadeInThread(),
|
||||
isObjectDetectingReady(false),
|
||||
shouldObjectDetectingResultsBeForgot(false),
|
||||
stateThread(STATE_THREAD_STOPPED),
|
||||
timeWhenDetectingThreadStartedWork(-1),
|
||||
parameters(params)
|
||||
{
|
||||
CV_Assert(_detector);
|
||||
|
||||
cascadeInThread = _detector;
|
||||
}
|
||||
|
||||
cv::DetectionBasedTracker::SeparateDetectionWork::~SeparateDetectionWork()
|
||||
{
|
||||
if(stateThread!=STATE_THREAD_STOPPED) {
|
||||
LOGE("\n\n\nATTENTION!!! dangerous algorithm error: destructor DetectionBasedTracker::DetectionBasedTracker::~SeparateDetectionWork is called before stopping the workthread");
|
||||
}
|
||||
second_workthread.join();
|
||||
}
|
||||
bool cv::DetectionBasedTracker::SeparateDetectionWork::run()
|
||||
{
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::run() --- start");
|
||||
std::unique_lock<std::mutex> mtx_lock(mtx);
|
||||
// unlocked when leaving scope
|
||||
if (stateThread != STATE_THREAD_STOPPED) {
|
||||
LOGE("DetectionBasedTracker::SeparateDetectionWork::run is called while the previous run is not stopped");
|
||||
return false;
|
||||
}
|
||||
stateThread=STATE_THREAD_WORKING_SLEEPING;
|
||||
second_workthread = std::thread(workcycleObjectDetectorFunction, (void*)this); //TODO: add attributes?
|
||||
objectDetectorThreadStartStop.wait(mtx_lock);
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::run --- end");
|
||||
return true;
|
||||
}
|
||||
|
||||
#define CATCH_ALL_AND_LOG(_block) \
|
||||
try { \
|
||||
_block; \
|
||||
} \
|
||||
catch(const cv::Exception& e) { \
|
||||
LOGE0("\n %s: ERROR: OpenCV Exception caught: \n'%s'\n\n", CV_Func, e.what()); \
|
||||
} catch(const std::exception& e) { \
|
||||
LOGE0("\n %s: ERROR: Exception caught: \n'%s'\n\n", CV_Func, e.what()); \
|
||||
} catch(...) { \
|
||||
LOGE0("\n %s: ERROR: UNKNOWN Exception caught\n\n", CV_Func); \
|
||||
}
|
||||
|
||||
void* cv::workcycleObjectDetectorFunction(void* p)
|
||||
{
|
||||
CATCH_ALL_AND_LOG({ ((cv::DetectionBasedTracker::SeparateDetectionWork*)p)->workcycleObjectDetector(); });
|
||||
try{
|
||||
((cv::DetectionBasedTracker::SeparateDetectionWork*)p)->init();
|
||||
} catch(...) {
|
||||
LOGE0("DetectionBasedTracker: workcycleObjectDetectorFunction: ERROR concerning pointer, received as the function parameter");
|
||||
}
|
||||
return NULL;
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector()
|
||||
{
|
||||
static double freq = getTickFrequency();
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- start");
|
||||
std::vector<Rect> objects;
|
||||
|
||||
CV_Assert(stateThread==STATE_THREAD_WORKING_SLEEPING);
|
||||
std::unique_lock<std::mutex> mtx_lock(mtx);
|
||||
{
|
||||
objectDetectorThreadStartStop.notify_one();
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- before waiting");
|
||||
CV_Assert(stateThread==STATE_THREAD_WORKING_SLEEPING);
|
||||
objectDetectorRun.wait(mtx_lock);
|
||||
if (isWorking()) {
|
||||
stateThread=STATE_THREAD_WORKING_WITH_IMAGE;
|
||||
}
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- after waiting");
|
||||
}
|
||||
mtx_lock.unlock();
|
||||
|
||||
bool isFirstStep=true;
|
||||
|
||||
isObjectDetectingReady=false;
|
||||
|
||||
while(isWorking())
|
||||
{
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- next step");
|
||||
|
||||
if (! isFirstStep) {
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- before waiting");
|
||||
CV_Assert(stateThread==STATE_THREAD_WORKING_SLEEPING);
|
||||
mtx_lock.lock();
|
||||
if (!isWorking()) {//it is a rare case, but may cause a crash
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- go out from the workcycle from inner part of lock just before waiting");
|
||||
mtx_lock.unlock();
|
||||
break;
|
||||
}
|
||||
CV_Assert(stateThread==STATE_THREAD_WORKING_SLEEPING);
|
||||
objectDetectorRun.wait(mtx_lock);
|
||||
if (isWorking()) {
|
||||
stateThread=STATE_THREAD_WORKING_WITH_IMAGE;
|
||||
}
|
||||
mtx_lock.unlock();
|
||||
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- after waiting");
|
||||
} else {
|
||||
isFirstStep=false;
|
||||
}
|
||||
|
||||
if (!isWorking()) {
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- go out from the workcycle just after waiting");
|
||||
break;
|
||||
}
|
||||
|
||||
|
||||
if (imageSeparateDetecting.empty()) {
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- imageSeparateDetecting is empty, continue");
|
||||
continue;
|
||||
}
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- start handling imageSeparateDetecting, img.size=%dx%d, img.data=0x%p",
|
||||
imageSeparateDetecting.size().width, imageSeparateDetecting.size().height, (void*)imageSeparateDetecting.data);
|
||||
|
||||
|
||||
int64 t1_detect=getTickCount();
|
||||
|
||||
cascadeInThread->detect(imageSeparateDetecting, objects);
|
||||
|
||||
/*cascadeInThread.detectMultiScale( imageSeparateDetecting, objects,
|
||||
detectionBasedTracker.parameters.scaleFactor, detectionBasedTracker.parameters.minNeighbors, 0
|
||||
|CV_HAAR_SCALE_IMAGE
|
||||
,
|
||||
min_objectSize,
|
||||
max_objectSize
|
||||
);
|
||||
*/
|
||||
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- end handling imageSeparateDetecting");
|
||||
|
||||
if (!isWorking()) {
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- go out from the workcycle just after detecting");
|
||||
break;
|
||||
}
|
||||
|
||||
int64 t2_detect = getTickCount();
|
||||
int64 dt_detect = t2_detect-t1_detect;
|
||||
double dt_detect_ms=((double)dt_detect)/freq * 1000.0;
|
||||
(void)(dt_detect_ms);
|
||||
|
||||
LOGI("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector() --- objects num==%d, t_ms=%.4f", (int)objects.size(), dt_detect_ms);
|
||||
mtx_lock.lock();
|
||||
if (!shouldObjectDetectingResultsBeForgot) {
|
||||
resultDetect=objects;
|
||||
isObjectDetectingReady=true;
|
||||
} else { //shouldObjectDetectingResultsBeForgot==true
|
||||
resultDetect.clear();
|
||||
isObjectDetectingReady=false;
|
||||
shouldObjectDetectingResultsBeForgot=false;
|
||||
}
|
||||
if(isWorking()) {
|
||||
stateThread=STATE_THREAD_WORKING_SLEEPING;
|
||||
}
|
||||
mtx_lock.unlock();
|
||||
|
||||
objects.clear();
|
||||
}// while(isWorking())
|
||||
|
||||
LOGI("DetectionBasedTracker::SeparateDetectionWork::workcycleObjectDetector: Returning");
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::SeparateDetectionWork::stop()
|
||||
{
|
||||
//FIXME: TODO: should add quickStop functionality
|
||||
std::unique_lock<std::mutex> mtx_lock(mtx);
|
||||
if (!isWorking()) {
|
||||
mtx_lock.unlock();
|
||||
LOGE("SimpleHighguiDemoCore::stop is called but the SimpleHighguiDemoCore pthread is not active");
|
||||
stateThread = STATE_THREAD_STOPPING;
|
||||
return;
|
||||
}
|
||||
stateThread=STATE_THREAD_STOPPING;
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::stop: before going to sleep to wait for the signal from the workthread");
|
||||
objectDetectorRun.notify_one();
|
||||
objectDetectorThreadStartStop.wait(mtx_lock);
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::stop: after receiving the signal from the workthread, stateThread=%d", (int)stateThread);
|
||||
mtx_lock.unlock();
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::SeparateDetectionWork::resetTracking()
|
||||
{
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::resetTracking");
|
||||
std::unique_lock<std::mutex> mtx_lock(mtx);
|
||||
|
||||
if (stateThread == STATE_THREAD_WORKING_WITH_IMAGE) {
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::resetTracking: since workthread is detecting objects at the moment, we should make cascadeInThread stop detecting and forget the detecting results");
|
||||
shouldObjectDetectingResultsBeForgot=true;
|
||||
//cascadeInThread.setStopFlag();//FIXME: TODO: this feature also should be contributed to OpenCV
|
||||
} else {
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::resetTracking: since workthread is NOT detecting objects at the moment, we should NOT make any additional actions");
|
||||
}
|
||||
|
||||
resultDetect.clear();
|
||||
isObjectDetectingReady=false;
|
||||
|
||||
mtx_lock.unlock();
|
||||
}
|
||||
|
||||
bool cv::DetectionBasedTracker::SeparateDetectionWork::communicateWithDetectingThread(const Mat& imageGray, std::vector<Rect>& rectsWhereRegions)
|
||||
{
|
||||
static double freq = getTickFrequency();
|
||||
|
||||
bool shouldCommunicateWithDetectingThread = (stateThread==STATE_THREAD_WORKING_SLEEPING);
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::communicateWithDetectingThread: shouldCommunicateWithDetectingThread=%d", (shouldCommunicateWithDetectingThread?1:0));
|
||||
|
||||
if (!shouldCommunicateWithDetectingThread) {
|
||||
return false;
|
||||
}
|
||||
|
||||
bool shouldHandleResult = false;
|
||||
|
||||
std::unique_lock<std::mutex> mtx_lock(mtx);
|
||||
|
||||
if (isObjectDetectingReady) {
|
||||
shouldHandleResult=true;
|
||||
rectsWhereRegions = resultDetect;
|
||||
isObjectDetectingReady=false;
|
||||
|
||||
double lastBigDetectionDuration = 1000.0 * (((double)(getTickCount() - timeWhenDetectingThreadStartedWork )) / freq);
|
||||
(void)(lastBigDetectionDuration);
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::communicateWithDetectingThread: lastBigDetectionDuration=%f ms", (double)lastBigDetectionDuration);
|
||||
}
|
||||
|
||||
bool shouldSendNewDataToWorkThread = true;
|
||||
if (timeWhenDetectingThreadStartedWork > 0) {
|
||||
double time_from_previous_launch_in_ms=1000.0 * (((double)(getTickCount() - timeWhenDetectingThreadStartedWork )) / freq); //the same formula as for lastBigDetectionDuration
|
||||
shouldSendNewDataToWorkThread = (time_from_previous_launch_in_ms >= detectionBasedTracker.parameters.minDetectionPeriod);
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::communicateWithDetectingThread: shouldSendNewDataToWorkThread was 1, now it is %d, since time_from_previous_launch_in_ms=%.2f, minDetectionPeriod=%d",
|
||||
(shouldSendNewDataToWorkThread?1:0), time_from_previous_launch_in_ms, detectionBasedTracker.parameters.minDetectionPeriod);
|
||||
}
|
||||
|
||||
if (shouldSendNewDataToWorkThread) {
|
||||
|
||||
imageSeparateDetecting.create(imageGray.size(), CV_8UC1);
|
||||
|
||||
imageGray.copyTo(imageSeparateDetecting);//may change imageSeparateDetecting ptr. But should not.
|
||||
|
||||
|
||||
timeWhenDetectingThreadStartedWork = getTickCount() ;
|
||||
|
||||
objectDetectorRun.notify_one();
|
||||
}
|
||||
|
||||
mtx_lock.unlock();
|
||||
LOGD("DetectionBasedTracker::SeparateDetectionWork::communicateWithDetectingThread: result: shouldHandleResult=%d", (shouldHandleResult?1:0));
|
||||
|
||||
return shouldHandleResult;
|
||||
}
|
||||
|
||||
cv::DetectionBasedTracker::Parameters::Parameters()
|
||||
{
|
||||
maxTrackLifetime = 5;
|
||||
minDetectionPeriod = 0;
|
||||
}
|
||||
|
||||
cv::DetectionBasedTracker::InnerParameters::InnerParameters()
|
||||
{
|
||||
numLastPositionsToTrack=4;
|
||||
numStepsToWaitBeforeFirstShow=6;
|
||||
numStepsToTrackWithoutDetectingIfObjectHasNotBeenShown=3;
|
||||
numStepsToShowWithoutDetecting=3;
|
||||
|
||||
coeffTrackingWindowSize=2.0;
|
||||
coeffObjectSizeToTrack=0.85f;
|
||||
coeffObjectSpeedUsingInPrediction=0.8f;
|
||||
|
||||
}
|
||||
|
||||
cv::DetectionBasedTracker::DetectionBasedTracker(cv::Ptr<IDetector> mainDetector, cv::Ptr<IDetector> trackingDetector, const Parameters& params)
|
||||
:separateDetectionWork(),
|
||||
parameters(params),
|
||||
innerParameters(),
|
||||
numTrackedSteps(0),
|
||||
cascadeForTracking(trackingDetector)
|
||||
{
|
||||
CV_Assert( (params.maxTrackLifetime >= 0)
|
||||
// && mainDetector
|
||||
&& trackingDetector );
|
||||
|
||||
if (mainDetector) {
|
||||
Ptr<SeparateDetectionWork> tmp(new SeparateDetectionWork(*this, mainDetector, params));
|
||||
separateDetectionWork.swap(tmp);
|
||||
}
|
||||
|
||||
weightsPositionsSmoothing.push_back(1);
|
||||
weightsSizesSmoothing.push_back(0.5);
|
||||
weightsSizesSmoothing.push_back(0.3f);
|
||||
weightsSizesSmoothing.push_back(0.2f);
|
||||
}
|
||||
|
||||
cv::DetectionBasedTracker::~DetectionBasedTracker()
|
||||
{
|
||||
}
|
||||
|
||||
void DetectionBasedTracker::process(const Mat& imageGray)
|
||||
{
|
||||
CV_INSTRUMENT_REGION();
|
||||
|
||||
CV_Assert(imageGray.type()==CV_8UC1);
|
||||
|
||||
if ( separateDetectionWork && !separateDetectionWork->isWorking() ) {
|
||||
separateDetectionWork->run();
|
||||
}
|
||||
|
||||
static double freq = getTickFrequency();
|
||||
static long long time_when_last_call_started=getTickCount();
|
||||
|
||||
{
|
||||
double delta_time_from_prev_call=1000.0 * (((double)(getTickCount() - time_when_last_call_started)) / freq);
|
||||
(void)(delta_time_from_prev_call);
|
||||
LOGD("DetectionBasedTracker::process: time from the previous call is %f ms", (double)delta_time_from_prev_call);
|
||||
time_when_last_call_started=getTickCount();
|
||||
}
|
||||
|
||||
Mat imageDetect=imageGray;
|
||||
|
||||
std::vector<Rect> rectsWhereRegions;
|
||||
bool shouldHandleResult=false;
|
||||
if (separateDetectionWork) {
|
||||
shouldHandleResult = separateDetectionWork->communicateWithDetectingThread(imageGray, rectsWhereRegions);
|
||||
}
|
||||
|
||||
if (shouldHandleResult) {
|
||||
LOGD("DetectionBasedTracker::process: get _rectsWhereRegions were got from resultDetect");
|
||||
} else {
|
||||
LOGD("DetectionBasedTracker::process: get _rectsWhereRegions from previous positions");
|
||||
for(size_t i = 0; i < trackedObjects.size(); i++) {
|
||||
size_t n = trackedObjects[i].lastPositions.size();
|
||||
CV_Assert(n > 0);
|
||||
|
||||
Rect r = trackedObjects[i].lastPositions[n-1];
|
||||
if(r.empty()) {
|
||||
LOGE("DetectionBasedTracker::process: ERROR: ATTENTION: strange algorithm's behavior: trackedObjects[i].rect() is empty");
|
||||
continue;
|
||||
}
|
||||
|
||||
//correction by speed of rectangle
|
||||
if (n > 1) {
|
||||
Point2f center = centerRect(r);
|
||||
Point2f center_prev = centerRect(trackedObjects[i].lastPositions[n-2]);
|
||||
Point2f shift = (center - center_prev) * innerParameters.coeffObjectSpeedUsingInPrediction;
|
||||
|
||||
r.x += cvRound(shift.x);
|
||||
r.y += cvRound(shift.y);
|
||||
}
|
||||
|
||||
|
||||
rectsWhereRegions.push_back(r);
|
||||
}
|
||||
}
|
||||
LOGI("DetectionBasedTracker::process: tracked objects num==%d", (int)trackedObjects.size());
|
||||
|
||||
std::vector<Rect> detectedObjectsInRegions;
|
||||
|
||||
LOGD("DetectionBasedTracker::process: rectsWhereRegions.size()=%d", (int)rectsWhereRegions.size());
|
||||
for(size_t i=0; i < rectsWhereRegions.size(); i++) {
|
||||
Rect r = rectsWhereRegions[i];
|
||||
|
||||
detectInRegion(imageDetect, r, detectedObjectsInRegions);
|
||||
}
|
||||
LOGD("DetectionBasedTracker::process: detectedObjectsInRegions.size()=%d", (int)detectedObjectsInRegions.size());
|
||||
|
||||
updateTrackedObjects(detectedObjectsInRegions);
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::getObjects(std::vector<cv::Rect>& result) const
|
||||
{
|
||||
result.clear();
|
||||
|
||||
for(size_t i=0; i < trackedObjects.size(); i++) {
|
||||
Rect r=calcTrackedObjectPositionToShow((int)i);
|
||||
if (r.empty()) {
|
||||
continue;
|
||||
}
|
||||
result.push_back(r);
|
||||
LOGD("DetectionBasedTracker::process: found a object with SIZE %d x %d, rect={%d, %d, %d x %d}", r.width, r.height, r.x, r.y, r.width, r.height);
|
||||
}
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::getObjects(std::vector<Object>& result) const
|
||||
{
|
||||
result.clear();
|
||||
|
||||
for(size_t i=0; i < trackedObjects.size(); i++) {
|
||||
Rect r=calcTrackedObjectPositionToShow((int)i);
|
||||
if (r.empty()) {
|
||||
continue;
|
||||
}
|
||||
result.push_back(Object(r, trackedObjects[i].id));
|
||||
LOGD("DetectionBasedTracker::process: found a object with SIZE %d x %d, rect={%d, %d, %d x %d}", r.width, r.height, r.x, r.y, r.width, r.height);
|
||||
}
|
||||
}
|
||||
void cv::DetectionBasedTracker::getObjects(std::vector<ExtObject>& result) const
|
||||
{
|
||||
result.clear();
|
||||
|
||||
for(size_t i=0; i < trackedObjects.size(); i++) {
|
||||
ObjectStatus status;
|
||||
Rect r=calcTrackedObjectPositionToShow((int)i, status);
|
||||
result.push_back(ExtObject(trackedObjects[i].id, r, status));
|
||||
LOGD("DetectionBasedTracker::process: found a object with SIZE %d x %d, rect={%d, %d, %d x %d}, status = %d", r.width, r.height, r.x, r.y, r.width, r.height, (int)status);
|
||||
}
|
||||
}
|
||||
|
||||
bool cv::DetectionBasedTracker::run()
|
||||
{
|
||||
if (separateDetectionWork) {
|
||||
return separateDetectionWork->run();
|
||||
}
|
||||
return false;
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::stop()
|
||||
{
|
||||
if (separateDetectionWork) {
|
||||
separateDetectionWork->stop();
|
||||
}
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::resetTracking()
|
||||
{
|
||||
if (separateDetectionWork) {
|
||||
separateDetectionWork->resetTracking();
|
||||
}
|
||||
trackedObjects.clear();
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::updateTrackedObjects(const std::vector<Rect>& detectedObjects)
|
||||
{
|
||||
enum {
|
||||
NEW_RECTANGLE=-1,
|
||||
INTERSECTED_RECTANGLE=-2
|
||||
};
|
||||
|
||||
int N1=(int)trackedObjects.size();
|
||||
int N2=(int)detectedObjects.size();
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: N1=%d, N2=%d", N1, N2);
|
||||
|
||||
for(int i=0; i < N1; i++) {
|
||||
trackedObjects[i].numDetectedFrames++;
|
||||
}
|
||||
|
||||
std::vector<int> correspondence(detectedObjects.size(), NEW_RECTANGLE);
|
||||
correspondence.clear();
|
||||
correspondence.resize(detectedObjects.size(), NEW_RECTANGLE);
|
||||
|
||||
for(int i=0; i < N1; i++) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: i=%d", i);
|
||||
TrackedObject& curObject=trackedObjects[i];
|
||||
|
||||
int bestIndex=-1;
|
||||
int bestArea=-1;
|
||||
|
||||
int numpositions=(int)curObject.lastPositions.size();
|
||||
CV_Assert(numpositions > 0);
|
||||
Rect prevRect=curObject.lastPositions[numpositions-1];
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: prevRect[%d]={%d, %d, %d x %d}", i, prevRect.x, prevRect.y, prevRect.width, prevRect.height);
|
||||
|
||||
for(int j=0; j < N2; j++) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: j=%d", j);
|
||||
if (correspondence[j] >= 0) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: j=%d is rejected, because it has correspondence=%d", j, correspondence[j]);
|
||||
continue;
|
||||
}
|
||||
if (correspondence[j] !=NEW_RECTANGLE) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: j=%d is rejected, because it is intersected with another rectangle", j);
|
||||
continue;
|
||||
}
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: detectedObjects[%d]={%d, %d, %d x %d}",
|
||||
j, detectedObjects[j].x, detectedObjects[j].y, detectedObjects[j].width, detectedObjects[j].height);
|
||||
|
||||
Rect r=prevRect & detectedObjects[j];
|
||||
if ( (r.width > 0) && (r.height > 0) ) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: There is intersection between prevRect and detectedRect, r={%d, %d, %d x %d}",
|
||||
r.x, r.y, r.width, r.height);
|
||||
correspondence[j]=INTERSECTED_RECTANGLE;
|
||||
|
||||
if ( r.area() > bestArea) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: The area of intersection is %d, it is better than bestArea=%d", r.area(), bestArea);
|
||||
bestIndex=j;
|
||||
bestArea=r.area();
|
||||
}
|
||||
}
|
||||
}
|
||||
if (bestIndex >= 0) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: The best correspondence for i=%d is j=%d", i, bestIndex);
|
||||
correspondence[bestIndex]=i;
|
||||
|
||||
for(int j=0; j < N2; j++) {
|
||||
if (correspondence[j] >= 0)
|
||||
continue;
|
||||
|
||||
Rect r=detectedObjects[j] & detectedObjects[bestIndex];
|
||||
if ( (r.width > 0) && (r.height > 0) ) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: Found intersection between "
|
||||
"rectangles j=%d and bestIndex=%d, rectangle j=%d is marked as intersected", j, bestIndex, j);
|
||||
correspondence[j]=INTERSECTED_RECTANGLE;
|
||||
}
|
||||
}
|
||||
} else {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: There is no correspondence for i=%d ", i);
|
||||
curObject.numFramesNotDetected++;
|
||||
}
|
||||
}
|
||||
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: start second cycle");
|
||||
for(int j=0; j < N2; j++) {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: j=%d", j);
|
||||
int i=correspondence[j];
|
||||
if (i >= 0) {//add position
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: add position");
|
||||
trackedObjects[i].lastPositions.push_back(detectedObjects[j]);
|
||||
while ((int)trackedObjects[i].lastPositions.size() > (int) innerParameters.numLastPositionsToTrack) {
|
||||
trackedObjects[i].lastPositions.erase(trackedObjects[i].lastPositions.begin());
|
||||
}
|
||||
trackedObjects[i].numFramesNotDetected=0;
|
||||
} else if (i==NEW_RECTANGLE){ //new object
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: new object");
|
||||
trackedObjects.push_back(detectedObjects[j]);
|
||||
} else {
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: was auxiliary intersection");
|
||||
}
|
||||
}
|
||||
|
||||
std::vector<TrackedObject>::iterator it=trackedObjects.begin();
|
||||
while( it != trackedObjects.end() ) {
|
||||
if ( (it->numFramesNotDetected > parameters.maxTrackLifetime)
|
||||
||
|
||||
(
|
||||
(it->numDetectedFrames <= innerParameters.numStepsToWaitBeforeFirstShow)
|
||||
&&
|
||||
(it->numFramesNotDetected > innerParameters.numStepsToTrackWithoutDetectingIfObjectHasNotBeenShown)
|
||||
)
|
||||
)
|
||||
{
|
||||
int numpos=(int)it->lastPositions.size();
|
||||
CV_Assert(numpos > 0);
|
||||
Rect r = it->lastPositions[numpos-1];
|
||||
(void)(r);
|
||||
LOGD("DetectionBasedTracker::updateTrackedObjects: deleted object {%d, %d, %d x %d}",
|
||||
r.x, r.y, r.width, r.height);
|
||||
it=trackedObjects.erase(it);
|
||||
} else {
|
||||
it++;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
int cv::DetectionBasedTracker::addObject(const Rect& location)
|
||||
{
|
||||
LOGD("DetectionBasedTracker::addObject: new object {%d, %d %dx%d}",location.x, location.y, location.width, location.height);
|
||||
trackedObjects.push_back(TrackedObject(location));
|
||||
int newId = trackedObjects.back().id;
|
||||
LOGD("DetectionBasedTracker::addObject: newId = %d", newId);
|
||||
return newId;
|
||||
}
|
||||
|
||||
Rect cv::DetectionBasedTracker::calcTrackedObjectPositionToShow(int i) const
|
||||
{
|
||||
ObjectStatus status;
|
||||
return calcTrackedObjectPositionToShow(i, status);
|
||||
}
|
||||
Rect cv::DetectionBasedTracker::calcTrackedObjectPositionToShow(int i, ObjectStatus& status) const
|
||||
{
|
||||
if ( (i < 0) || (i >= (int)trackedObjects.size()) ) {
|
||||
LOGE("DetectionBasedTracker::calcTrackedObjectPositionToShow: ERROR: wrong i=%d", i);
|
||||
status = WRONG_OBJECT;
|
||||
return Rect();
|
||||
}
|
||||
if (trackedObjects[i].numDetectedFrames <= innerParameters.numStepsToWaitBeforeFirstShow){
|
||||
LOGI("DetectionBasedTracker::calcTrackedObjectPositionToShow: trackedObjects[%d].numDetectedFrames=%d <= numStepsToWaitBeforeFirstShow=%d --- return empty Rect()",
|
||||
i, trackedObjects[i].numDetectedFrames, innerParameters.numStepsToWaitBeforeFirstShow);
|
||||
status = DETECTED_NOT_SHOWN_YET;
|
||||
return Rect();
|
||||
}
|
||||
if (trackedObjects[i].numFramesNotDetected > innerParameters.numStepsToShowWithoutDetecting) {
|
||||
status = DETECTED_TEMPORARY_LOST;
|
||||
return Rect();
|
||||
}
|
||||
|
||||
const TrackedObject::PositionsVector& lastPositions=trackedObjects[i].lastPositions;
|
||||
|
||||
int N=(int)lastPositions.size();
|
||||
if (N<=0) {
|
||||
LOGE("DetectionBasedTracker::calcTrackedObjectPositionToShow: ERROR: no positions for i=%d", i);
|
||||
status = WRONG_OBJECT;
|
||||
return Rect();
|
||||
}
|
||||
|
||||
int Nsize=std::min(N, (int)weightsSizesSmoothing.size());
|
||||
int Ncenter= std::min(N, (int)weightsPositionsSmoothing.size());
|
||||
|
||||
Point2f center;
|
||||
double w=0, h=0;
|
||||
if (Nsize > 0) {
|
||||
double sum=0;
|
||||
for(int j=0; j < Nsize; j++) {
|
||||
int k=N-j-1;
|
||||
w += lastPositions[k].width * weightsSizesSmoothing[j];
|
||||
h += lastPositions[k].height * weightsSizesSmoothing[j];
|
||||
sum+=weightsSizesSmoothing[j];
|
||||
}
|
||||
w /= sum;
|
||||
h /= sum;
|
||||
} else {
|
||||
w=lastPositions[N-1].width;
|
||||
h=lastPositions[N-1].height;
|
||||
}
|
||||
|
||||
if (Ncenter > 0) {
|
||||
double sum=0;
|
||||
for(int j=0; j < Ncenter; j++) {
|
||||
int k=N-j-1;
|
||||
Point tl(lastPositions[k].tl());
|
||||
Point br(lastPositions[k].br());
|
||||
Point2f c1;
|
||||
c1=tl;
|
||||
c1=c1* 0.5f;
|
||||
Point2f c2;
|
||||
c2=br;
|
||||
c2=c2*0.5f;
|
||||
c1=c1+c2;
|
||||
|
||||
center=center+ (c1 * weightsPositionsSmoothing[j]);
|
||||
sum+=weightsPositionsSmoothing[j];
|
||||
}
|
||||
center *= (float)(1 / sum);
|
||||
} else {
|
||||
int k=N-1;
|
||||
Point tl(lastPositions[k].tl());
|
||||
Point br(lastPositions[k].br());
|
||||
Point2f c1;
|
||||
c1=tl;
|
||||
c1=c1* 0.5f;
|
||||
Point2f c2;
|
||||
c2=br;
|
||||
c2=c2*0.5f;
|
||||
|
||||
center=c1+c2;
|
||||
}
|
||||
Point2f tl=center-Point2f((float)w*0.5f,(float)h*0.5f);
|
||||
Rect res(cvRound(tl.x), cvRound(tl.y), cvRound(w), cvRound(h));
|
||||
LOGD("DetectionBasedTracker::calcTrackedObjectPositionToShow: Result for i=%d: {%d, %d, %d x %d}", i, res.x, res.y, res.width, res.height);
|
||||
|
||||
status = DETECTED;
|
||||
return res;
|
||||
}
|
||||
|
||||
void cv::DetectionBasedTracker::detectInRegion(const Mat& img, const Rect& r, std::vector<Rect>& detectedObjectsInRegions)
|
||||
{
|
||||
Rect r0(Point(), img.size());
|
||||
Rect r1 = scale_rect(r, innerParameters.coeffTrackingWindowSize);
|
||||
r1 = r1 & r0;
|
||||
|
||||
if ( (r1.width <=0) || (r1.height <= 0) ) {
|
||||
LOGD("DetectionBasedTracker::detectInRegion: Empty intersection");
|
||||
return;
|
||||
}
|
||||
|
||||
int d = cvRound(std::min(r.width, r.height) * innerParameters.coeffObjectSizeToTrack);
|
||||
|
||||
std::vector<Rect> tmpobjects;
|
||||
|
||||
Mat img1(img, r1);//subimage for rectangle -- without data copying
|
||||
LOGD("DetectionBasedTracker::detectInRegion: img1.size()=%d x %d, d=%d",
|
||||
img1.size().width, img1.size().height, d);
|
||||
|
||||
cascadeForTracking->setMinObjectSize(Size(d, d));
|
||||
cascadeForTracking->detect(img1, tmpobjects);
|
||||
/*
|
||||
detectMultiScale( img1, tmpobjects,
|
||||
parameters.scaleFactor, parameters.minNeighbors, 0
|
||||
|CV_HAAR_FIND_BIGGEST_OBJECT
|
||||
|CV_HAAR_SCALE_IMAGE
|
||||
,
|
||||
Size(d,d),
|
||||
max_objectSize
|
||||
);*/
|
||||
|
||||
for(size_t i=0; i < tmpobjects.size(); i++) {
|
||||
Rect curres(tmpobjects[i].tl() + r1.tl(), tmpobjects[i].size());
|
||||
detectedObjectsInRegions.push_back(curres);
|
||||
}
|
||||
}
|
||||
|
||||
bool cv::DetectionBasedTracker::setParameters(const Parameters& params)
|
||||
{
|
||||
if ( params.maxTrackLifetime < 0 )
|
||||
{
|
||||
LOGE("DetectionBasedTracker::setParameters: ERROR: wrong parameters value");
|
||||
return false;
|
||||
}
|
||||
|
||||
if (separateDetectionWork) {
|
||||
separateDetectionWork->setParameters(params);
|
||||
}
|
||||
parameters=params;
|
||||
return true;
|
||||
}
|
||||
|
||||
const cv::DetectionBasedTracker::Parameters& DetectionBasedTracker::getParameters() const
|
||||
{
|
||||
return parameters;
|
||||
}
|
||||
3565
Lib/opencv/sources/modules/objdetect/src/hog.cpp
Normal file
3565
Lib/opencv/sources/modules/objdetect/src/hog.cpp
Normal file
File diff suppressed because it is too large
Load Diff
52
Lib/opencv/sources/modules/objdetect/src/main.cpp
Normal file
52
Lib/opencv/sources/modules/objdetect/src/main.cpp
Normal file
@@ -0,0 +1,52 @@
|
||||
/*M///////////////////////////////////////////////////////////////////////////////////////
|
||||
//
|
||||
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
|
||||
//
|
||||
// By downloading, copying, installing or using the software you agree to this license.
|
||||
// If you do not agree to this license, do not download, install,
|
||||
// copy or use the software.
|
||||
//
|
||||
//
|
||||
// License Agreement
|
||||
// For Open Source Computer Vision Library
|
||||
//
|
||||
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
|
||||
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
|
||||
// Copyright (C) 2015, Itseez Inc., all rights reserved.
|
||||
// Third party copyrights are property of their respective owners.
|
||||
//
|
||||
// Redistribution and use in source and binary forms, with or without modification,
|
||||
// are permitted provided that the following conditions are met:
|
||||
//
|
||||
// * Redistribution's of source code must retain the above copyright notice,
|
||||
// this list of conditions and the following disclaimer.
|
||||
//
|
||||
// * Redistribution's in binary form must reproduce the above copyright notice,
|
||||
// this list of conditions and the following disclaimer in the documentation
|
||||
// and/or other materials provided with the distribution.
|
||||
//
|
||||
// * The name of the copyright holders may not be used to endorse or promote products
|
||||
// derived from this software without specific prior written permission.
|
||||
//
|
||||
// This software is provided by the copyright holders and contributors "as is" and
|
||||
// any express or implied warranties, including, but not limited to, the implied
|
||||
// warranties of merchantability and fitness for a particular purpose are disclaimed.
|
||||
// In no event shall the Intel Corporation or contributors be liable for any direct,
|
||||
// indirect, incidental, special, exemplary, or consequential damages
|
||||
// (including, but not limited to, procurement of substitute goods or services;
|
||||
// loss of use, data, or profits; or business interruption) however caused
|
||||
// and on any theory of liability, whether in contract, strict liability,
|
||||
// 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.
|
||||
//
|
||||
//M*/
|
||||
|
||||
//
|
||||
// Library initialization file
|
||||
//
|
||||
|
||||
#include "precomp.hpp"
|
||||
|
||||
IPP_INITIALIZER_AUTO
|
||||
|
||||
/* End of file. */
|
||||
661
Lib/opencv/sources/modules/objdetect/src/opencl/cascadedetect.cl
Normal file
661
Lib/opencv/sources/modules/objdetect/src/opencl/cascadedetect.cl
Normal file
@@ -0,0 +1,661 @@
|
||||
///////////////////////////// OpenCL kernels for face detection //////////////////////////////
|
||||
////////////////////////////// see the opencv/doc/license.txt ///////////////////////////////
|
||||
|
||||
//
|
||||
// the code has been derived from the OpenCL Haar cascade kernel by
|
||||
//
|
||||
// Niko Li, newlife20080214@gmail.com
|
||||
// Wang Weiyan, wangweiyanster@gmail.com
|
||||
// Jia Haipeng, jiahaipeng95@gmail.com
|
||||
// Nathan, liujun@multicorewareinc.com
|
||||
// Peng Xiao, pengxiao@outlook.com
|
||||
// Erping Pang, erping@multicorewareinc.com
|
||||
//
|
||||
|
||||
#ifdef HAAR
|
||||
typedef struct __attribute__((aligned(4))) OptHaarFeature
|
||||
{
|
||||
int4 ofs[3] __attribute__((aligned (4)));
|
||||
float4 weight __attribute__((aligned (4)));
|
||||
}
|
||||
OptHaarFeature;
|
||||
#endif
|
||||
|
||||
#ifdef LBP
|
||||
typedef struct __attribute__((aligned(4))) OptLBPFeature
|
||||
{
|
||||
int16 ofs __attribute__((aligned (4)));
|
||||
}
|
||||
OptLBPFeature;
|
||||
#endif
|
||||
|
||||
typedef struct __attribute__((aligned(4))) Stump
|
||||
{
|
||||
float4 st __attribute__((aligned (4)));
|
||||
}
|
||||
Stump;
|
||||
|
||||
typedef struct __attribute__((aligned(4))) Node
|
||||
{
|
||||
int4 n __attribute__((aligned (4)));
|
||||
}
|
||||
Node;
|
||||
|
||||
typedef struct __attribute__((aligned (4))) Stage
|
||||
{
|
||||
int first __attribute__((aligned (4)));
|
||||
int ntrees __attribute__((aligned (4)));
|
||||
float threshold __attribute__((aligned (4)));
|
||||
}
|
||||
Stage;
|
||||
|
||||
typedef struct __attribute__((aligned (4))) ScaleData
|
||||
{
|
||||
float scale __attribute__((aligned (4)));
|
||||
int szi_width __attribute__((aligned (4)));
|
||||
int szi_height __attribute__((aligned (4)));
|
||||
int layer_ofs __attribute__((aligned (4)));
|
||||
int ystep __attribute__((aligned (4)));
|
||||
}
|
||||
ScaleData;
|
||||
|
||||
#ifndef SUM_BUF_SIZE
|
||||
#define SUM_BUF_SIZE 0
|
||||
#endif
|
||||
|
||||
#ifndef NODE_COUNT
|
||||
#define NODE_COUNT 1
|
||||
#endif
|
||||
|
||||
#ifdef HAAR
|
||||
__kernel __attribute__((reqd_work_group_size(LOCAL_SIZE_X,LOCAL_SIZE_Y,1)))
|
||||
void runHaarClassifier(
|
||||
int nscales, __global const ScaleData* scaleData,
|
||||
__global const int* sum,
|
||||
int _sumstep, int sumoffset,
|
||||
__global const OptHaarFeature* optfeatures,
|
||||
__global const Stage* stages,
|
||||
__global const Node* nodes,
|
||||
__global const float* leaves0,
|
||||
|
||||
volatile __global int* facepos,
|
||||
int4 normrect, int sqofs, int2 windowsize)
|
||||
{
|
||||
int lx = get_local_id(0);
|
||||
int ly = get_local_id(1);
|
||||
int groupIdx = get_group_id(0);
|
||||
int i, ngroups = get_global_size(0)/LOCAL_SIZE_X;
|
||||
int scaleIdx, tileIdx, stageIdx;
|
||||
int sumstep = (int)(_sumstep/sizeof(int));
|
||||
int4 nofs0 = (int4)(mad24(normrect.y, sumstep, normrect.x),
|
||||
mad24(normrect.y, sumstep, normrect.x + normrect.z),
|
||||
mad24(normrect.y + normrect.w, sumstep, normrect.x),
|
||||
mad24(normrect.y + normrect.w, sumstep, normrect.x + normrect.z));
|
||||
int normarea = normrect.z * normrect.w;
|
||||
float invarea = 1.f/normarea;
|
||||
int lidx = ly*LOCAL_SIZE_X + lx;
|
||||
|
||||
#if SUM_BUF_SIZE > 0
|
||||
int4 nofs = (int4)(mad24(normrect.y, SUM_BUF_STEP, normrect.x),
|
||||
mad24(normrect.y, SUM_BUF_STEP, normrect.x + normrect.z),
|
||||
mad24(normrect.y + normrect.w, SUM_BUF_STEP, normrect.x),
|
||||
mad24(normrect.y + normrect.w, SUM_BUF_STEP, normrect.x + normrect.z));
|
||||
#else
|
||||
int4 nofs = nofs0;
|
||||
#endif
|
||||
#define LOCAL_SIZE (LOCAL_SIZE_X*LOCAL_SIZE_Y)
|
||||
__local int lstore[SUM_BUF_SIZE + LOCAL_SIZE*5/2+1];
|
||||
#if SUM_BUF_SIZE > 0
|
||||
__local int* ibuf = lstore;
|
||||
__local int* lcount = ibuf + SUM_BUF_SIZE;
|
||||
#else
|
||||
__local int* lcount = lstore;
|
||||
#endif
|
||||
__local float* lnf = (__local float*)(lcount + 1);
|
||||
__local float* lpartsum = lnf + LOCAL_SIZE;
|
||||
__local short* lbuf = (__local short*)(lpartsum + LOCAL_SIZE);
|
||||
|
||||
for( scaleIdx = nscales-1; scaleIdx >= 0; scaleIdx-- )
|
||||
{
|
||||
__global const ScaleData* s = scaleData + scaleIdx;
|
||||
int ystep = s->ystep;
|
||||
int2 worksize = (int2)(max(s->szi_width - windowsize.x, 0), max(s->szi_height - windowsize.y, 0));
|
||||
int2 ntiles = (int2)((worksize.x + LOCAL_SIZE_X-1)/LOCAL_SIZE_X,
|
||||
(worksize.y + LOCAL_SIZE_Y-1)/LOCAL_SIZE_Y);
|
||||
int totalTiles = ntiles.x*ntiles.y;
|
||||
|
||||
for( tileIdx = groupIdx; tileIdx < totalTiles; tileIdx += ngroups )
|
||||
{
|
||||
int ix0 = (tileIdx % ntiles.x)*LOCAL_SIZE_X;
|
||||
int iy0 = (tileIdx / ntiles.x)*LOCAL_SIZE_Y;
|
||||
int ix = lx, iy = ly;
|
||||
__global const int* psum0 = sum + mad24(iy0, sumstep, ix0) + s->layer_ofs;
|
||||
__global const int* psum1 = psum0 + mad24(iy, sumstep, ix);
|
||||
|
||||
if( ix0 >= worksize.x || iy0 >= worksize.y )
|
||||
continue;
|
||||
#if SUM_BUF_SIZE > 0
|
||||
for( i = lidx*4; i < SUM_BUF_SIZE; i += LOCAL_SIZE_X*LOCAL_SIZE_Y*4 )
|
||||
{
|
||||
int dy = i/SUM_BUF_STEP, dx = i - dy*SUM_BUF_STEP;
|
||||
vstore4(vload4(0, psum0 + mad24(dy, sumstep, dx)), 0, ibuf+i);
|
||||
}
|
||||
#endif
|
||||
|
||||
if( lidx == 0 )
|
||||
lcount[0] = 0;
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if( ix0 + ix < worksize.x && iy0 + iy < worksize.y )
|
||||
{
|
||||
#if NODE_COUNT==1
|
||||
__global const Stump* stump = (__global const Stump*)nodes;
|
||||
#else
|
||||
__global const Node* node = nodes;
|
||||
__global const float* leaves = leaves0;
|
||||
#endif
|
||||
#if SUM_BUF_SIZE > 0
|
||||
__local const int* psum = ibuf + mad24(iy, SUM_BUF_STEP, ix);
|
||||
#else
|
||||
__global const int* psum = psum1;
|
||||
#endif
|
||||
|
||||
__global const int* psqsum = (__global const int*)(psum1 + sqofs);
|
||||
float sval = (psum[nofs.x] - psum[nofs.y] - psum[nofs.z] + psum[nofs.w])*invarea;
|
||||
float sqval = (psqsum[nofs0.x] - psqsum[nofs0.y] - psqsum[nofs0.z] + psqsum[nofs0.w])*invarea;
|
||||
float nf = (float)normarea * sqrt(max(sqval - sval * sval, 0.f));
|
||||
nf = nf > 0 ? nf : 1.f;
|
||||
|
||||
for( stageIdx = 0; stageIdx < SPLIT_STAGE; stageIdx++ )
|
||||
{
|
||||
int ntrees = stages[stageIdx].ntrees;
|
||||
float s = 0.f;
|
||||
#if NODE_COUNT==1
|
||||
for( i = 0; i < ntrees; i++ )
|
||||
{
|
||||
float4 st = stump[i].st;
|
||||
__global const OptHaarFeature* f = optfeatures + as_int(st.x);
|
||||
float4 weight = f->weight;
|
||||
|
||||
int4 ofs = f->ofs[0];
|
||||
sval = (psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w])*weight.x;
|
||||
ofs = f->ofs[1];
|
||||
sval = mad((float)(psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.y, sval);
|
||||
if( weight.z > 0 )
|
||||
{
|
||||
ofs = f->ofs[2];
|
||||
sval = mad((float)(psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.z, sval);
|
||||
}
|
||||
|
||||
s += (sval < st.y*nf) ? st.z : st.w;
|
||||
}
|
||||
stump += ntrees;
|
||||
#else
|
||||
for( i = 0; i < ntrees; i++, node += NODE_COUNT, leaves += NODE_COUNT+1 )
|
||||
{
|
||||
int idx = 0;
|
||||
do
|
||||
{
|
||||
int4 n = node[idx].n;
|
||||
__global const OptHaarFeature* f = optfeatures + n.x;
|
||||
float4 weight = f->weight;
|
||||
|
||||
int4 ofs = f->ofs[0];
|
||||
|
||||
sval = (psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w])*weight.x;
|
||||
ofs = f->ofs[1];
|
||||
sval = mad((float)(psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.y, sval);
|
||||
if( weight.z > 0 )
|
||||
{
|
||||
ofs = f->ofs[2];
|
||||
sval = mad((float)(psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.z, sval);
|
||||
}
|
||||
|
||||
idx = (sval < as_float(n.y)*nf) ? n.z : n.w;
|
||||
}
|
||||
while(idx > 0);
|
||||
s += leaves[-idx];
|
||||
}
|
||||
#endif
|
||||
|
||||
if( s < stages[stageIdx].threshold )
|
||||
break;
|
||||
}
|
||||
|
||||
if( stageIdx == SPLIT_STAGE && (ystep == 1 || ((ix | iy) & 1) == 0) )
|
||||
{
|
||||
int count = atomic_inc(lcount);
|
||||
lbuf[count] = (int)(ix | (iy << 8));
|
||||
lnf[count] = nf;
|
||||
}
|
||||
}
|
||||
|
||||
for( stageIdx = SPLIT_STAGE; stageIdx < N_STAGES; stageIdx++ )
|
||||
{
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
int nrects = lcount[0];
|
||||
|
||||
if( nrects == 0 )
|
||||
break;
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if( lidx == 0 )
|
||||
lcount[0] = 0;
|
||||
|
||||
{
|
||||
#if NODE_COUNT == 1
|
||||
__global const Stump* stump = (__global const Stump*)nodes + stages[stageIdx].first;
|
||||
#else
|
||||
__global const Node* node = nodes + stages[stageIdx].first*NODE_COUNT;
|
||||
__global const float* leaves = leaves0 + stages[stageIdx].first*(NODE_COUNT+1);
|
||||
#endif
|
||||
int nparts = LOCAL_SIZE / nrects;
|
||||
int ntrees = stages[stageIdx].ntrees;
|
||||
int ntrees_p = (ntrees + nparts - 1)/nparts;
|
||||
int nr = lidx / nparts;
|
||||
int partidx = -1, idxval = 0;
|
||||
float partsum = 0.f, nf = 0.f;
|
||||
|
||||
if( nr < nrects )
|
||||
{
|
||||
partidx = lidx % nparts;
|
||||
idxval = lbuf[nr];
|
||||
nf = lnf[nr];
|
||||
|
||||
{
|
||||
int ntrees0 = ntrees_p*partidx;
|
||||
int ntrees1 = min(ntrees0 + ntrees_p, ntrees);
|
||||
int ix1 = idxval & 255, iy1 = idxval >> 8;
|
||||
#if SUM_BUF_SIZE > 0
|
||||
__local const int* psum = ibuf + mad24(iy1, SUM_BUF_STEP, ix1);
|
||||
#else
|
||||
__global const int* psum = psum0 + mad24(iy1, sumstep, ix1);
|
||||
#endif
|
||||
|
||||
#if NODE_COUNT == 1
|
||||
for( i = ntrees0; i < ntrees1; i++ )
|
||||
{
|
||||
float4 st = stump[i].st;
|
||||
__global const OptHaarFeature* f = optfeatures + as_int(st.x);
|
||||
float4 weight = f->weight;
|
||||
|
||||
int4 ofs = f->ofs[0];
|
||||
float sval = (psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w])*weight.x;
|
||||
ofs = f->ofs[1];
|
||||
sval = mad((float)(psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.y, sval);
|
||||
//if( weight.z > 0 )
|
||||
if( fabs(weight.z) > 0 )
|
||||
{
|
||||
ofs = f->ofs[2];
|
||||
sval = mad((float)(psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.z, sval);
|
||||
}
|
||||
|
||||
partsum += (sval < st.y*nf) ? st.z : st.w;
|
||||
}
|
||||
#else
|
||||
for( i = ntrees0; i < ntrees1; i++ )
|
||||
{
|
||||
int idx = 0;
|
||||
do
|
||||
{
|
||||
int4 n = node[i*2 + idx].n;
|
||||
__global const OptHaarFeature* f = optfeatures + n.x;
|
||||
float4 weight = f->weight;
|
||||
int4 ofs = f->ofs[0];
|
||||
|
||||
float sval = (psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w])*weight.x;
|
||||
ofs = f->ofs[1];
|
||||
sval = mad((float)(psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.y, sval);
|
||||
if( weight.z > 0 )
|
||||
{
|
||||
ofs = f->ofs[2];
|
||||
sval = mad((float)(psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.z, sval);
|
||||
}
|
||||
|
||||
idx = (sval < as_float(n.y)*nf) ? n.z : n.w;
|
||||
}
|
||||
while(idx > 0);
|
||||
partsum += leaves[i*3-idx];
|
||||
}
|
||||
#endif
|
||||
}
|
||||
}
|
||||
lpartsum[lidx] = partsum;
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if( partidx == 0 )
|
||||
{
|
||||
float s = lpartsum[nr*nparts];
|
||||
for( i = 1; i < nparts; i++ )
|
||||
s += lpartsum[i + nr*nparts];
|
||||
if( s >= stages[stageIdx].threshold )
|
||||
{
|
||||
int count = atomic_inc(lcount);
|
||||
lbuf[count] = idxval;
|
||||
lnf[count] = nf;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if( stageIdx == N_STAGES )
|
||||
{
|
||||
int nrects = lcount[0];
|
||||
if( lidx < nrects )
|
||||
{
|
||||
int nfaces = atomic_inc(facepos);
|
||||
if( nfaces < MAX_FACES )
|
||||
{
|
||||
volatile __global int* face = facepos + 1 + nfaces*3;
|
||||
int val = lbuf[lidx];
|
||||
face[0] = scaleIdx;
|
||||
face[1] = ix0 + (val & 255);
|
||||
face[2] = iy0 + (val >> 8);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif
|
||||
|
||||
#ifdef LBP
|
||||
#undef CALC_SUM_OFS_
|
||||
#define CALC_SUM_OFS_(p0, p1, p2, p3, ptr) \
|
||||
((ptr)[p0] - (ptr)[p1] - (ptr)[p2] + (ptr)[p3])
|
||||
|
||||
__kernel void runLBPClassifierStumpSimple(
|
||||
int nscales, __global const ScaleData* scaleData,
|
||||
__global const int* sum,
|
||||
int _sumstep, int sumoffset,
|
||||
__global const OptLBPFeature* optfeatures,
|
||||
__global const Stage* stages,
|
||||
__global const Stump* stumps,
|
||||
__global const int* bitsets,
|
||||
int bitsetSize,
|
||||
|
||||
volatile __global int* facepos,
|
||||
int2 windowsize)
|
||||
{
|
||||
int lx = get_local_id(0);
|
||||
int ly = get_local_id(1);
|
||||
int local_size_x = get_local_size(0);
|
||||
int local_size_y = get_local_size(1);
|
||||
int groupIdx = get_group_id(1)*get_num_groups(0) + get_group_id(0);
|
||||
int ngroups = get_num_groups(0)*get_num_groups(1);
|
||||
int scaleIdx, tileIdx, stageIdx;
|
||||
int sumstep = (int)(_sumstep/sizeof(int));
|
||||
|
||||
for( scaleIdx = nscales-1; scaleIdx >= 0; scaleIdx-- )
|
||||
{
|
||||
__global const ScaleData* s = scaleData + scaleIdx;
|
||||
int ystep = s->ystep;
|
||||
int2 worksize = (int2)(max(s->szi_width - windowsize.x, 0), max(s->szi_height - windowsize.y, 0));
|
||||
int2 ntiles = (int2)((worksize.x/ystep + local_size_x-1)/local_size_x,
|
||||
(worksize.y/ystep + local_size_y-1)/local_size_y);
|
||||
int totalTiles = ntiles.x*ntiles.y;
|
||||
|
||||
for( tileIdx = groupIdx; tileIdx < totalTiles; tileIdx += ngroups )
|
||||
{
|
||||
int iy = mad24((tileIdx / ntiles.x), local_size_y, ly) * ystep;
|
||||
int ix = mad24((tileIdx % ntiles.x), local_size_x, lx) * ystep;
|
||||
|
||||
if( ix < worksize.x && iy < worksize.y )
|
||||
{
|
||||
__global const int* p = sum + mad24(iy, sumstep, ix) + s->layer_ofs;
|
||||
__global const Stump* stump = stumps;
|
||||
__global const int* bitset = bitsets;
|
||||
|
||||
for( stageIdx = 0; stageIdx < N_STAGES; stageIdx++ )
|
||||
{
|
||||
int i, ntrees = stages[stageIdx].ntrees;
|
||||
float s = 0.f;
|
||||
for( i = 0; i < ntrees; i++, stump++, bitset += bitsetSize )
|
||||
{
|
||||
float4 st = stump->st;
|
||||
__global const OptLBPFeature* f = optfeatures + as_int(st.x);
|
||||
int16 ofs = f->ofs;
|
||||
|
||||
int cval = CALC_SUM_OFS_( ofs.s5, ofs.s6, ofs.s9, ofs.sa, p );
|
||||
|
||||
int mask, idx = (CALC_SUM_OFS_( ofs.s0, ofs.s1, ofs.s4, ofs.s5, p ) >= cval ? 4 : 0); // 0
|
||||
idx |= (CALC_SUM_OFS_( ofs.s1, ofs.s2, ofs.s5, ofs.s6, p ) >= cval ? 2 : 0); // 1
|
||||
idx |= (CALC_SUM_OFS_( ofs.s2, ofs.s3, ofs.s6, ofs.s7, p ) >= cval ? 1 : 0); // 2
|
||||
|
||||
mask = (CALC_SUM_OFS_( ofs.s6, ofs.s7, ofs.sa, ofs.sb, p ) >= cval ? 16 : 0); // 5
|
||||
mask |= (CALC_SUM_OFS_( ofs.sa, ofs.sb, ofs.se, ofs.sf, p ) >= cval ? 8 : 0); // 8
|
||||
mask |= (CALC_SUM_OFS_( ofs.s9, ofs.sa, ofs.sd, ofs.se, p ) >= cval ? 4 : 0); // 7
|
||||
mask |= (CALC_SUM_OFS_( ofs.s8, ofs.s9, ofs.sc, ofs.sd, p ) >= cval ? 2 : 0); // 6
|
||||
mask |= (CALC_SUM_OFS_( ofs.s4, ofs.s5, ofs.s8, ofs.s9, p ) >= cval ? 1 : 0); // 7
|
||||
|
||||
s += (bitset[idx] & (1 << mask)) ? st.z : st.w;
|
||||
}
|
||||
|
||||
if( s < stages[stageIdx].threshold )
|
||||
break;
|
||||
}
|
||||
|
||||
if( stageIdx == N_STAGES )
|
||||
{
|
||||
int nfaces = atomic_inc(facepos);
|
||||
if( nfaces < MAX_FACES )
|
||||
{
|
||||
volatile __global int* face = facepos + 1 + nfaces*3;
|
||||
face[0] = scaleIdx;
|
||||
face[1] = ix;
|
||||
face[2] = iy;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
__kernel __attribute__((reqd_work_group_size(LOCAL_SIZE_X,LOCAL_SIZE_Y,1)))
|
||||
void runLBPClassifierStump(
|
||||
int nscales, __global const ScaleData* scaleData,
|
||||
__global const int* sum,
|
||||
int _sumstep, int sumoffset,
|
||||
__global const OptLBPFeature* optfeatures,
|
||||
__global const Stage* stages,
|
||||
__global const Stump* stumps,
|
||||
__global const int* bitsets,
|
||||
int bitsetSize,
|
||||
|
||||
volatile __global int* facepos,
|
||||
int2 windowsize)
|
||||
{
|
||||
int lx = get_local_id(0);
|
||||
int ly = get_local_id(1);
|
||||
int groupIdx = get_group_id(0);
|
||||
int i, ngroups = get_global_size(0)/LOCAL_SIZE_X;
|
||||
int scaleIdx, tileIdx, stageIdx;
|
||||
int sumstep = (int)(_sumstep/sizeof(int));
|
||||
int lidx = ly*LOCAL_SIZE_X + lx;
|
||||
|
||||
#define LOCAL_SIZE (LOCAL_SIZE_X*LOCAL_SIZE_Y)
|
||||
__local int lstore[SUM_BUF_SIZE + LOCAL_SIZE*3/2+1];
|
||||
#if SUM_BUF_SIZE > 0
|
||||
__local int* ibuf = lstore;
|
||||
__local int* lcount = ibuf + SUM_BUF_SIZE;
|
||||
#else
|
||||
__local int* lcount = lstore;
|
||||
#endif
|
||||
__local float* lpartsum = (__local float*)(lcount + 1);
|
||||
__local short* lbuf = (__local short*)(lpartsum + LOCAL_SIZE);
|
||||
|
||||
for( scaleIdx = nscales-1; scaleIdx >= 0; scaleIdx-- )
|
||||
{
|
||||
__global const ScaleData* s = scaleData + scaleIdx;
|
||||
int ystep = s->ystep;
|
||||
int2 worksize = (int2)(max(s->szi_width - windowsize.x, 0), max(s->szi_height - windowsize.y, 0));
|
||||
int2 ntiles = (int2)((worksize.x + LOCAL_SIZE_X-1)/LOCAL_SIZE_X,
|
||||
(worksize.y + LOCAL_SIZE_Y-1)/LOCAL_SIZE_Y);
|
||||
int totalTiles = ntiles.x*ntiles.y;
|
||||
|
||||
for( tileIdx = groupIdx; tileIdx < totalTiles; tileIdx += ngroups )
|
||||
{
|
||||
int ix0 = (tileIdx % ntiles.x)*LOCAL_SIZE_X;
|
||||
int iy0 = (tileIdx / ntiles.x)*LOCAL_SIZE_Y;
|
||||
int ix = lx, iy = ly;
|
||||
__global const int* psum0 = sum + mad24(iy0, sumstep, ix0) + s->layer_ofs;
|
||||
|
||||
if( ix0 >= worksize.x || iy0 >= worksize.y )
|
||||
continue;
|
||||
#if SUM_BUF_SIZE > 0
|
||||
for( i = lidx*4; i < SUM_BUF_SIZE; i += LOCAL_SIZE_X*LOCAL_SIZE_Y*4 )
|
||||
{
|
||||
int dy = i/SUM_BUF_STEP, dx = i - dy*SUM_BUF_STEP;
|
||||
vstore4(vload4(0, psum0 + mad24(dy, sumstep, dx)), 0, ibuf+i);
|
||||
}
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
#endif
|
||||
|
||||
if( lidx == 0 )
|
||||
lcount[0] = 0;
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if( ix0 + ix < worksize.x && iy0 + iy < worksize.y )
|
||||
{
|
||||
__global const Stump* stump = stumps;
|
||||
__global const int* bitset = bitsets;
|
||||
#if SUM_BUF_SIZE > 0
|
||||
__local const int* p = ibuf + mad24(iy, SUM_BUF_STEP, ix);
|
||||
#else
|
||||
__global const int* p = psum0 + mad24(iy, sumstep, ix);
|
||||
#endif
|
||||
|
||||
for( stageIdx = 0; stageIdx < SPLIT_STAGE; stageIdx++ )
|
||||
{
|
||||
int ntrees = stages[stageIdx].ntrees;
|
||||
float s = 0.f;
|
||||
for( i = 0; i < ntrees; i++, stump++, bitset += bitsetSize )
|
||||
{
|
||||
float4 st = stump->st;
|
||||
__global const OptLBPFeature* f = optfeatures + as_int(st.x);
|
||||
int16 ofs = f->ofs;
|
||||
|
||||
int cval = CALC_SUM_OFS_( ofs.s5, ofs.s6, ofs.s9, ofs.sa, p );
|
||||
|
||||
int mask, idx = (CALC_SUM_OFS_( ofs.s0, ofs.s1, ofs.s4, ofs.s5, p ) >= cval ? 4 : 0); // 0
|
||||
idx |= (CALC_SUM_OFS_( ofs.s1, ofs.s2, ofs.s5, ofs.s6, p ) >= cval ? 2 : 0); // 1
|
||||
idx |= (CALC_SUM_OFS_( ofs.s2, ofs.s3, ofs.s6, ofs.s7, p ) >= cval ? 1 : 0); // 2
|
||||
|
||||
mask = (CALC_SUM_OFS_( ofs.s6, ofs.s7, ofs.sa, ofs.sb, p ) >= cval ? 16 : 0); // 5
|
||||
mask |= (CALC_SUM_OFS_( ofs.sa, ofs.sb, ofs.se, ofs.sf, p ) >= cval ? 8 : 0); // 8
|
||||
mask |= (CALC_SUM_OFS_( ofs.s9, ofs.sa, ofs.sd, ofs.se, p ) >= cval ? 4 : 0); // 7
|
||||
mask |= (CALC_SUM_OFS_( ofs.s8, ofs.s9, ofs.sc, ofs.sd, p ) >= cval ? 2 : 0); // 6
|
||||
mask |= (CALC_SUM_OFS_( ofs.s4, ofs.s5, ofs.s8, ofs.s9, p ) >= cval ? 1 : 0); // 7
|
||||
|
||||
s += (bitset[idx] & (1 << mask)) ? st.z : st.w;
|
||||
}
|
||||
|
||||
if( s < stages[stageIdx].threshold )
|
||||
break;
|
||||
}
|
||||
|
||||
if( stageIdx == SPLIT_STAGE && (ystep == 1 || ((ix | iy) & 1) == 0) )
|
||||
{
|
||||
int count = atomic_inc(lcount);
|
||||
lbuf[count] = (int)(ix | (iy << 8));
|
||||
}
|
||||
}
|
||||
|
||||
for( stageIdx = SPLIT_STAGE; stageIdx < N_STAGES; stageIdx++ )
|
||||
{
|
||||
int nrects = lcount[0];
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if( nrects == 0 )
|
||||
break;
|
||||
if( lidx == 0 )
|
||||
lcount[0] = 0;
|
||||
|
||||
{
|
||||
__global const Stump* stump = stumps + stages[stageIdx].first;
|
||||
__global const int* bitset = bitsets + stages[stageIdx].first*bitsetSize;
|
||||
int nparts = LOCAL_SIZE / nrects;
|
||||
int ntrees = stages[stageIdx].ntrees;
|
||||
int ntrees_p = (ntrees + nparts - 1)/nparts;
|
||||
int nr = lidx / nparts;
|
||||
int partidx = -1, idxval = 0;
|
||||
float partsum = 0.f, nf = 0.f;
|
||||
|
||||
if( nr < nrects )
|
||||
{
|
||||
partidx = lidx % nparts;
|
||||
idxval = lbuf[nr];
|
||||
|
||||
{
|
||||
int ntrees0 = ntrees_p*partidx;
|
||||
int ntrees1 = min(ntrees0 + ntrees_p, ntrees);
|
||||
int ix1 = idxval & 255, iy1 = idxval >> 8;
|
||||
#if SUM_BUF_SIZE > 0
|
||||
__local const int* p = ibuf + mad24(iy1, SUM_BUF_STEP, ix1);
|
||||
#else
|
||||
__global const int* p = psum0 + mad24(iy1, sumstep, ix1);
|
||||
#endif
|
||||
|
||||
for( i = ntrees0; i < ntrees1; i++ )
|
||||
{
|
||||
float4 st = stump[i].st;
|
||||
__global const OptLBPFeature* f = optfeatures + as_int(st.x);
|
||||
int16 ofs = f->ofs;
|
||||
|
||||
#define CALC_SUM_OFS_(p0, p1, p2, p3, ptr) \
|
||||
((ptr)[p0] - (ptr)[p1] - (ptr)[p2] + (ptr)[p3])
|
||||
|
||||
int cval = CALC_SUM_OFS_( ofs.s5, ofs.s6, ofs.s9, ofs.sa, p );
|
||||
|
||||
int mask, idx = (CALC_SUM_OFS_( ofs.s0, ofs.s1, ofs.s4, ofs.s5, p ) >= cval ? 4 : 0); // 0
|
||||
idx |= (CALC_SUM_OFS_( ofs.s1, ofs.s2, ofs.s5, ofs.s6, p ) >= cval ? 2 : 0); // 1
|
||||
idx |= (CALC_SUM_OFS_( ofs.s2, ofs.s3, ofs.s6, ofs.s7, p ) >= cval ? 1 : 0); // 2
|
||||
|
||||
mask = (CALC_SUM_OFS_( ofs.s6, ofs.s7, ofs.sa, ofs.sb, p ) >= cval ? 16 : 0); // 5
|
||||
mask |= (CALC_SUM_OFS_( ofs.sa, ofs.sb, ofs.se, ofs.sf, p ) >= cval ? 8 : 0); // 8
|
||||
mask |= (CALC_SUM_OFS_( ofs.s9, ofs.sa, ofs.sd, ofs.se, p ) >= cval ? 4 : 0); // 7
|
||||
mask |= (CALC_SUM_OFS_( ofs.s8, ofs.s9, ofs.sc, ofs.sd, p ) >= cval ? 2 : 0); // 6
|
||||
mask |= (CALC_SUM_OFS_( ofs.s4, ofs.s5, ofs.s8, ofs.s9, p ) >= cval ? 1 : 0); // 7
|
||||
|
||||
partsum += (bitset[i*bitsetSize + idx] & (1 << mask)) ? st.z : st.w;
|
||||
}
|
||||
}
|
||||
}
|
||||
lpartsum[lidx] = partsum;
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if( partidx == 0 )
|
||||
{
|
||||
float s = lpartsum[nr*nparts];
|
||||
for( i = 1; i < nparts; i++ )
|
||||
s += lpartsum[i + nr*nparts];
|
||||
if( s >= stages[stageIdx].threshold )
|
||||
{
|
||||
int count = atomic_inc(lcount);
|
||||
lbuf[count] = idxval;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if( stageIdx == N_STAGES )
|
||||
{
|
||||
int nrects = lcount[0];
|
||||
if( lidx < nrects )
|
||||
{
|
||||
int nfaces = atomic_inc(facepos);
|
||||
if( nfaces < MAX_FACES )
|
||||
{
|
||||
volatile __global int* face = facepos + 1 + nfaces*3;
|
||||
int val = lbuf[lidx];
|
||||
face[0] = scaleIdx;
|
||||
face[1] = ix0 + (val & 255);
|
||||
face[2] = iy0 + (val >> 8);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif
|
||||
722
Lib/opencv/sources/modules/objdetect/src/opencl/objdetect_hog.cl
Normal file
722
Lib/opencv/sources/modules/objdetect/src/opencl/objdetect_hog.cl
Normal file
@@ -0,0 +1,722 @@
|
||||
/*M///////////////////////////////////////////////////////////////////////////////////////
|
||||
//
|
||||
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
|
||||
//
|
||||
// By downloading, copying, installing or using the software you agree to this license.
|
||||
// If you do not agree to this license, do not download, install,
|
||||
// copy or use the software.
|
||||
//
|
||||
//
|
||||
// License Agreement
|
||||
// For Open Source Computer Vision Library
|
||||
//
|
||||
// Copyright (C) 2010-2012, Multicoreware, Inc., all rights reserved.
|
||||
// Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
|
||||
// Third party copyrights are property of their respective owners.
|
||||
//
|
||||
// @Authors
|
||||
// Wenju He, wenju@multicorewareinc.com
|
||||
//
|
||||
// Redistribution and use in source and binary forms, with or without modification,
|
||||
// are permitted provided that the following conditions are met:
|
||||
//
|
||||
// * Redistribution's of source code must retain the above copyright notice,
|
||||
// this list of conditions and the following disclaimer.
|
||||
//
|
||||
// * Redistribution's in binary form must reproduce the above copyright notice,
|
||||
// this list of conditions and the following disclaimer in the documentation
|
||||
// and/or other materials provided with the distribution.
|
||||
//
|
||||
// * The name of the copyright holders may not be used to endorse or promote products
|
||||
// derived from this software without specific prior written permission.
|
||||
//
|
||||
// This software is provided by the copyright holders and contributors as is and
|
||||
// any express or implied warranties, including, but not limited to, the implied
|
||||
// warranties of merchantability and fitness for a particular purpose are disclaimed.
|
||||
// In no event shall the Intel Corporation or contributors be liable for any direct,
|
||||
// indirect, incidental, special, exemplary, or consequential damages
|
||||
// (including, but not limited to, procurement of substitute goods or services;
|
||||
// loss of use, data, or profits; or business interruption) however caused
|
||||
// and on any theory of liability, whether in contract, strict liability,
|
||||
// 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.
|
||||
//
|
||||
//M*/
|
||||
|
||||
#define CELL_WIDTH 8
|
||||
#define CELL_HEIGHT 8
|
||||
#define CELLS_PER_BLOCK_X 2
|
||||
#define CELLS_PER_BLOCK_Y 2
|
||||
#define NTHREADS 256
|
||||
#define CV_PI_F M_PI_F
|
||||
|
||||
#ifdef INTEL_DEVICE
|
||||
#define QANGLE_TYPE int
|
||||
#define QANGLE_TYPE2 int2
|
||||
#else
|
||||
#define QANGLE_TYPE uchar
|
||||
#define QANGLE_TYPE2 uchar2
|
||||
#endif
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Histogram computation
|
||||
// 12 threads for a cell, 12x4 threads per block
|
||||
// Use pre-computed gaussian and interp_weight lookup tables
|
||||
__kernel void compute_hists_lut_kernel(
|
||||
const int cblock_stride_x, const int cblock_stride_y,
|
||||
const int cnbins, const int cblock_hist_size, const int img_block_width,
|
||||
const int blocks_in_group, const int blocks_total,
|
||||
const int grad_quadstep, const int qangle_step,
|
||||
__global const float* grad, __global const QANGLE_TYPE* qangle,
|
||||
__global const float* gauss_w_lut,
|
||||
__global float* block_hists, __local float* smem)
|
||||
{
|
||||
const int lx = get_local_id(0);
|
||||
const int lp = lx / 24; /* local group id */
|
||||
const int gid = get_group_id(0) * blocks_in_group + lp;/* global group id */
|
||||
const int gidY = gid / img_block_width;
|
||||
const int gidX = gid - gidY * img_block_width;
|
||||
|
||||
const int lidX = lx - lp * 24;
|
||||
const int lidY = get_local_id(1);
|
||||
|
||||
const int cell_x = lidX / 12;
|
||||
const int cell_y = lidY;
|
||||
const int cell_thread_x = lidX - cell_x * 12;
|
||||
|
||||
__local float* hists = smem + lp * cnbins * (CELLS_PER_BLOCK_X *
|
||||
CELLS_PER_BLOCK_Y * 12 + CELLS_PER_BLOCK_X * CELLS_PER_BLOCK_Y);
|
||||
__local float* final_hist = hists + cnbins *
|
||||
(CELLS_PER_BLOCK_X * CELLS_PER_BLOCK_Y * 12);
|
||||
|
||||
const int offset_x = gidX * cblock_stride_x + (cell_x << 2) + cell_thread_x;
|
||||
const int offset_y = gidY * cblock_stride_y + (cell_y << 2);
|
||||
|
||||
__global const float* grad_ptr = (gid < blocks_total) ?
|
||||
grad + offset_y * grad_quadstep + (offset_x << 1) : grad;
|
||||
__global const QANGLE_TYPE* qangle_ptr = (gid < blocks_total) ?
|
||||
qangle + offset_y * qangle_step + (offset_x << 1) : qangle;
|
||||
|
||||
__local float* hist = hists + 12 * (cell_y * CELLS_PER_BLOCK_Y + cell_x) +
|
||||
cell_thread_x;
|
||||
for (int bin_id = 0; bin_id < cnbins; ++bin_id)
|
||||
hist[bin_id * 48] = 0.f;
|
||||
|
||||
const int dist_x = -4 + cell_thread_x - 4 * cell_x;
|
||||
const int dist_center_x = dist_x - 4 * (1 - 2 * cell_x);
|
||||
|
||||
const int dist_y_begin = -4 - 4 * lidY;
|
||||
for (int dist_y = dist_y_begin; dist_y < dist_y_begin + 12; ++dist_y)
|
||||
{
|
||||
float2 vote = (float2) (grad_ptr[0], grad_ptr[1]);
|
||||
QANGLE_TYPE2 bin = (QANGLE_TYPE2) (qangle_ptr[0], qangle_ptr[1]);
|
||||
|
||||
grad_ptr += grad_quadstep;
|
||||
qangle_ptr += qangle_step;
|
||||
|
||||
int dist_center_y = dist_y - 4 * (1 - 2 * cell_y);
|
||||
|
||||
int idx = (dist_center_y + 8) * 16 + (dist_center_x + 8);
|
||||
float gaussian = gauss_w_lut[idx];
|
||||
idx = (dist_y + 8) * 16 + (dist_x + 8);
|
||||
float interp_weight = gauss_w_lut[256+idx];
|
||||
|
||||
hist[bin.x * 48] += gaussian * interp_weight * vote.x;
|
||||
hist[bin.y * 48] += gaussian * interp_weight * vote.y;
|
||||
}
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
volatile __local float* hist_ = hist;
|
||||
for (int bin_id = 0; bin_id < cnbins; ++bin_id, hist_ += 48)
|
||||
{
|
||||
if (cell_thread_x < 6)
|
||||
hist_[0] += hist_[6];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (cell_thread_x < 3)
|
||||
hist_[0] += hist_[3];
|
||||
#ifdef CPU
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
#endif
|
||||
if (cell_thread_x == 0)
|
||||
final_hist[(cell_x * 2 + cell_y) * cnbins + bin_id] =
|
||||
hist_[0] + hist_[1] + hist_[2];
|
||||
}
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
int tid = (cell_y * CELLS_PER_BLOCK_Y + cell_x) * 12 + cell_thread_x;
|
||||
if ((tid < cblock_hist_size) && (gid < blocks_total))
|
||||
{
|
||||
__global float* block_hist = block_hists +
|
||||
(gidY * img_block_width + gidX) * cblock_hist_size;
|
||||
block_hist[tid] = final_hist[tid];
|
||||
}
|
||||
}
|
||||
|
||||
//-------------------------------------------------------------
|
||||
// Normalization of histograms via L2Hys_norm
|
||||
// optimized for the case of 9 bins
|
||||
__kernel void normalize_hists_36_kernel(__global float* block_hists,
|
||||
const float threshold, __local float *squares)
|
||||
{
|
||||
const int tid = get_local_id(0);
|
||||
const int gid = get_global_id(0);
|
||||
const int bid = tid / 36; /* block-hist id, (0 - 6) */
|
||||
const int boffset = bid * 36; /* block-hist offset in the work-group */
|
||||
const int hid = tid - boffset; /* histogram bin id, (0 - 35) */
|
||||
|
||||
float elem = block_hists[gid];
|
||||
squares[tid] = elem * elem;
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
__local float* smem = squares + boffset;
|
||||
float sum = smem[hid];
|
||||
if (hid < 18)
|
||||
smem[hid] = sum = sum + smem[hid + 18];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (hid < 9)
|
||||
smem[hid] = sum = sum + smem[hid + 9];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (hid < 4)
|
||||
smem[hid] = sum + smem[hid + 4];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
sum = smem[0] + smem[1] + smem[2] + smem[3] + smem[8];
|
||||
|
||||
elem = elem / (sqrt(sum) + 3.6f);
|
||||
elem = min(elem, threshold);
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
squares[tid] = elem * elem;
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
sum = smem[hid];
|
||||
if (hid < 18)
|
||||
smem[hid] = sum = sum + smem[hid + 18];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (hid < 9)
|
||||
smem[hid] = sum = sum + smem[hid + 9];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (hid < 4)
|
||||
smem[hid] = sum + smem[hid + 4];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
sum = smem[0] + smem[1] + smem[2] + smem[3] + smem[8];
|
||||
|
||||
block_hists[gid] = elem / (sqrt(sum) + 1e-3f);
|
||||
}
|
||||
|
||||
//-------------------------------------------------------------
|
||||
// Normalization of histograms via L2Hys_norm
|
||||
//
|
||||
inline float reduce_smem(volatile __local float* smem, int size)
|
||||
{
|
||||
unsigned int tid = get_local_id(0);
|
||||
float sum = smem[tid];
|
||||
|
||||
if (size >= 512) { if (tid < 256) smem[tid] = sum = sum + smem[tid + 256];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
if (size >= 256) { if (tid < 128) smem[tid] = sum = sum + smem[tid + 128];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
if (size >= 128) { if (tid < 64) smem[tid] = sum = sum + smem[tid + 64];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
#ifdef CPU
|
||||
if (size >= 64) { if (tid < 32) smem[tid] = sum = sum + smem[tid + 32];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
if (size >= 32) { if (tid < 16) smem[tid] = sum = sum + smem[tid + 16];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
if (size >= 16) { if (tid < 8) smem[tid] = sum = sum + smem[tid + 8];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
if (size >= 8) { if (tid < 4) smem[tid] = sum = sum + smem[tid + 4];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
if (size >= 4) { if (tid < 2) smem[tid] = sum = sum + smem[tid + 2];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
if (size >= 2) { if (tid < 1) smem[tid] = sum = sum + smem[tid + 1];
|
||||
barrier(CLK_LOCAL_MEM_FENCE); }
|
||||
#else
|
||||
if (tid < 32)
|
||||
{
|
||||
if (size >= 64) smem[tid] = sum = sum + smem[tid + 32];
|
||||
#if WAVE_SIZE < 32
|
||||
} barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (tid < 16) {
|
||||
#endif
|
||||
if (size >= 32) smem[tid] = sum = sum + smem[tid + 16];
|
||||
if (size >= 16) smem[tid] = sum = sum + smem[tid + 8];
|
||||
if (size >= 8) smem[tid] = sum = sum + smem[tid + 4];
|
||||
if (size >= 4) smem[tid] = sum = sum + smem[tid + 2];
|
||||
if (size >= 2) smem[tid] = sum = sum + smem[tid + 1];
|
||||
}
|
||||
#endif
|
||||
|
||||
return sum;
|
||||
}
|
||||
|
||||
__kernel void normalize_hists_kernel(
|
||||
const int nthreads, const int block_hist_size, const int img_block_width,
|
||||
__global float* block_hists, const float threshold, __local float *squares)
|
||||
{
|
||||
const int tid = get_local_id(0);
|
||||
const int gidX = get_group_id(0);
|
||||
const int gidY = get_group_id(1);
|
||||
|
||||
__global float* hist = block_hists + (gidY * img_block_width + gidX) *
|
||||
block_hist_size + tid;
|
||||
|
||||
float elem = 0.f;
|
||||
if (tid < block_hist_size)
|
||||
elem = hist[0];
|
||||
|
||||
squares[tid] = elem * elem;
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
float sum = reduce_smem(squares, nthreads);
|
||||
|
||||
float scale = 1.0f / (sqrt(sum) + 0.1f * block_hist_size);
|
||||
elem = min(elem * scale, threshold);
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
squares[tid] = elem * elem;
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
sum = reduce_smem(squares, nthreads);
|
||||
scale = 1.0f / (sqrt(sum) + 1e-3f);
|
||||
|
||||
if (tid < block_hist_size)
|
||||
hist[0] = elem * scale;
|
||||
}
|
||||
|
||||
//---------------------------------------------------------------------
|
||||
// Linear SVM based classification
|
||||
// 48x96 window, 9 bins and default parameters
|
||||
// 180 threads, each thread corresponds to a bin in a row
|
||||
__kernel void classify_hists_180_kernel(
|
||||
const int cdescr_width, const int cdescr_height, const int cblock_hist_size,
|
||||
const int img_win_width, const int img_block_width,
|
||||
const int win_block_stride_x, const int win_block_stride_y,
|
||||
__global const float * block_hists, __global const float* coefs,
|
||||
float free_coef, float threshold, __global uchar* labels)
|
||||
{
|
||||
const int tid = get_local_id(0);
|
||||
const int gidX = get_group_id(0);
|
||||
const int gidY = get_group_id(1);
|
||||
|
||||
__global const float* hist = block_hists + (gidY * win_block_stride_y *
|
||||
img_block_width + gidX * win_block_stride_x) * cblock_hist_size;
|
||||
|
||||
float product = 0.f;
|
||||
|
||||
for (int i = 0; i < cdescr_height; i++)
|
||||
{
|
||||
product += coefs[i * cdescr_width + tid] *
|
||||
hist[i * img_block_width * cblock_hist_size + tid];
|
||||
}
|
||||
|
||||
__local float products[180];
|
||||
|
||||
products[tid] = product;
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if (tid < 90) products[tid] = product = product + products[tid + 90];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if (tid < 45) products[tid] = product = product + products[tid + 45];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
volatile __local float* smem = products;
|
||||
#ifdef CPU
|
||||
if (tid < 13) smem[tid] = product = product + smem[tid + 32];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (tid < 16) smem[tid] = product = product + smem[tid + 16];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<8) smem[tid] = product = product + smem[tid + 8];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<4) smem[tid] = product = product + smem[tid + 4];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<2) smem[tid] = product = product + smem[tid + 2];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
#else
|
||||
if (tid < 13)
|
||||
{
|
||||
smem[tid] = product = product + smem[tid + 32];
|
||||
}
|
||||
#if WAVE_SIZE < 32
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
#endif
|
||||
if (tid < 16)
|
||||
{
|
||||
smem[tid] = product = product + smem[tid + 16];
|
||||
smem[tid] = product = product + smem[tid + 8];
|
||||
smem[tid] = product = product + smem[tid + 4];
|
||||
smem[tid] = product = product + smem[tid + 2];
|
||||
}
|
||||
#endif
|
||||
|
||||
if (tid == 0){
|
||||
product = product + smem[tid + 1];
|
||||
labels[gidY * img_win_width + gidX] = (product + free_coef >= threshold);
|
||||
}
|
||||
}
|
||||
|
||||
//---------------------------------------------------------------------
|
||||
// Linear SVM based classification
|
||||
// 64x128 window, 9 bins and default parameters
|
||||
// 256 threads, 252 of them are used
|
||||
__kernel void classify_hists_252_kernel(
|
||||
const int cdescr_width, const int cdescr_height, const int cblock_hist_size,
|
||||
const int img_win_width, const int img_block_width,
|
||||
const int win_block_stride_x, const int win_block_stride_y,
|
||||
__global const float * block_hists, __global const float* coefs,
|
||||
float free_coef, float threshold, __global uchar* labels)
|
||||
{
|
||||
const int tid = get_local_id(0);
|
||||
const int gidX = get_group_id(0);
|
||||
const int gidY = get_group_id(1);
|
||||
|
||||
__global const float* hist = block_hists + (gidY * win_block_stride_y *
|
||||
img_block_width + gidX * win_block_stride_x) * cblock_hist_size;
|
||||
|
||||
float product = 0.f;
|
||||
if (tid < cdescr_width)
|
||||
{
|
||||
for (int i = 0; i < cdescr_height; i++)
|
||||
product += coefs[i * cdescr_width + tid] *
|
||||
hist[i * img_block_width * cblock_hist_size + tid];
|
||||
}
|
||||
|
||||
__local float products[NTHREADS];
|
||||
|
||||
products[tid] = product;
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if (tid < 128) products[tid] = product = product + products[tid + 128];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if (tid < 64) products[tid] = product = product + products[tid + 64];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
volatile __local float* smem = products;
|
||||
#ifdef CPU
|
||||
if(tid<32) smem[tid] = product = product + smem[tid + 32];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<16) smem[tid] = product = product + smem[tid + 16];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<8) smem[tid] = product = product + smem[tid + 8];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<4) smem[tid] = product = product + smem[tid + 4];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<2) smem[tid] = product = product + smem[tid + 2];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
#else
|
||||
if (tid < 32)
|
||||
{
|
||||
smem[tid] = product = product + smem[tid + 32];
|
||||
#if WAVE_SIZE < 32
|
||||
} barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (tid < 16) {
|
||||
#endif
|
||||
smem[tid] = product = product + smem[tid + 16];
|
||||
smem[tid] = product = product + smem[tid + 8];
|
||||
smem[tid] = product = product + smem[tid + 4];
|
||||
smem[tid] = product = product + smem[tid + 2];
|
||||
}
|
||||
#endif
|
||||
if (tid == 0){
|
||||
product = product + smem[tid + 1];
|
||||
labels[gidY * img_win_width + gidX] = (product + free_coef >= threshold);
|
||||
}
|
||||
}
|
||||
|
||||
//---------------------------------------------------------------------
|
||||
// Linear SVM based classification
|
||||
// 256 threads
|
||||
__kernel void classify_hists_kernel(
|
||||
const int cdescr_size, const int cdescr_width, const int cblock_hist_size,
|
||||
const int img_win_width, const int img_block_width,
|
||||
const int win_block_stride_x, const int win_block_stride_y,
|
||||
__global const float * block_hists, __global const float* coefs,
|
||||
float free_coef, float threshold, __global uchar* labels)
|
||||
{
|
||||
const int tid = get_local_id(0);
|
||||
const int gidX = get_group_id(0);
|
||||
const int gidY = get_group_id(1);
|
||||
|
||||
__global const float* hist = block_hists + (gidY * win_block_stride_y *
|
||||
img_block_width + gidX * win_block_stride_x) * cblock_hist_size;
|
||||
|
||||
float product = 0.f;
|
||||
for (int i = tid; i < cdescr_size; i += NTHREADS)
|
||||
{
|
||||
int offset_y = i / cdescr_width;
|
||||
int offset_x = i - offset_y * cdescr_width;
|
||||
product += coefs[i] *
|
||||
hist[offset_y * img_block_width * cblock_hist_size + offset_x];
|
||||
}
|
||||
|
||||
__local float products[NTHREADS];
|
||||
|
||||
products[tid] = product;
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if (tid < 128) products[tid] = product = product + products[tid + 128];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
if (tid < 64) products[tid] = product = product + products[tid + 64];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
|
||||
volatile __local float* smem = products;
|
||||
#ifdef CPU
|
||||
if(tid<32) smem[tid] = product = product + smem[tid + 32];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<16) smem[tid] = product = product + smem[tid + 16];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<8) smem[tid] = product = product + smem[tid + 8];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<4) smem[tid] = product = product + smem[tid + 4];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if(tid<2) smem[tid] = product = product + smem[tid + 2];
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
#else
|
||||
if (tid < 32)
|
||||
{
|
||||
smem[tid] = product = product + smem[tid + 32];
|
||||
#if WAVE_SIZE < 32
|
||||
} barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (tid < 16) {
|
||||
#endif
|
||||
smem[tid] = product = product + smem[tid + 16];
|
||||
smem[tid] = product = product + smem[tid + 8];
|
||||
smem[tid] = product = product + smem[tid + 4];
|
||||
smem[tid] = product = product + smem[tid + 2];
|
||||
}
|
||||
#endif
|
||||
if (tid == 0){
|
||||
smem[tid] = product = product + smem[tid + 1];
|
||||
labels[gidY * img_win_width + gidX] = (product + free_coef >= threshold);
|
||||
}
|
||||
}
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Extract descriptors
|
||||
|
||||
__kernel void extract_descrs_by_rows_kernel(
|
||||
const int cblock_hist_size, const int descriptors_quadstep,
|
||||
const int cdescr_size, const int cdescr_width, const int img_block_width,
|
||||
const int win_block_stride_x, const int win_block_stride_y,
|
||||
__global const float* block_hists, __global float* descriptors)
|
||||
{
|
||||
int tid = get_local_id(0);
|
||||
int gidX = get_group_id(0);
|
||||
int gidY = get_group_id(1);
|
||||
|
||||
// Get left top corner of the window in src
|
||||
__global const float* hist = block_hists + (gidY * win_block_stride_y *
|
||||
img_block_width + gidX * win_block_stride_x) * cblock_hist_size;
|
||||
|
||||
// Get left top corner of the window in dst
|
||||
__global float* descriptor = descriptors +
|
||||
(gidY * get_num_groups(0) + gidX) * descriptors_quadstep;
|
||||
|
||||
// Copy elements from src to dst
|
||||
for (int i = tid; i < cdescr_size; i += NTHREADS)
|
||||
{
|
||||
int offset_y = i / cdescr_width;
|
||||
int offset_x = i - offset_y * cdescr_width;
|
||||
descriptor[i] = hist[offset_y * img_block_width * cblock_hist_size + offset_x];
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void extract_descrs_by_cols_kernel(
|
||||
const int cblock_hist_size, const int descriptors_quadstep, const int cdescr_size,
|
||||
const int cnblocks_win_x, const int cnblocks_win_y, const int img_block_width,
|
||||
const int win_block_stride_x, const int win_block_stride_y,
|
||||
__global const float* block_hists, __global float* descriptors)
|
||||
{
|
||||
int tid = get_local_id(0);
|
||||
int gidX = get_group_id(0);
|
||||
int gidY = get_group_id(1);
|
||||
|
||||
// Get left top corner of the window in src
|
||||
__global const float* hist = block_hists + (gidY * win_block_stride_y *
|
||||
img_block_width + gidX * win_block_stride_x) * cblock_hist_size;
|
||||
|
||||
// Get left top corner of the window in dst
|
||||
__global float* descriptor = descriptors +
|
||||
(gidY * get_num_groups(0) + gidX) * descriptors_quadstep;
|
||||
|
||||
// Copy elements from src to dst
|
||||
for (int i = tid; i < cdescr_size; i += NTHREADS)
|
||||
{
|
||||
int block_idx = i / cblock_hist_size;
|
||||
int idx_in_block = i - block_idx * cblock_hist_size;
|
||||
|
||||
int y = block_idx / cnblocks_win_x;
|
||||
int x = block_idx - y * cnblocks_win_x;
|
||||
|
||||
descriptor[(x * cnblocks_win_y + y) * cblock_hist_size + idx_in_block] =
|
||||
hist[(y * img_block_width + x) * cblock_hist_size + idx_in_block];
|
||||
}
|
||||
}
|
||||
|
||||
//----------------------------------------------------------------------------
|
||||
// Gradients computation
|
||||
|
||||
__kernel void compute_gradients_8UC4_kernel(
|
||||
const int height, const int width,
|
||||
const int img_step, const int grad_quadstep, const int qangle_step,
|
||||
const __global uchar4 * img, __global float * grad, __global QANGLE_TYPE * qangle,
|
||||
const float angle_scale, const char correct_gamma, const int cnbins)
|
||||
{
|
||||
const int x = get_global_id(0);
|
||||
const int tid = get_local_id(0);
|
||||
const int gSizeX = get_local_size(0);
|
||||
const int gidY = get_group_id(1);
|
||||
|
||||
__global const uchar4* row = img + gidY * img_step;
|
||||
|
||||
__local float sh_row[(NTHREADS + 2) * 3];
|
||||
|
||||
uchar4 val;
|
||||
if (x < width)
|
||||
val = row[x];
|
||||
else
|
||||
val = row[width - 2];
|
||||
|
||||
sh_row[tid + 1] = val.x;
|
||||
sh_row[tid + 1 + (NTHREADS + 2)] = val.y;
|
||||
sh_row[tid + 1 + 2 * (NTHREADS + 2)] = val.z;
|
||||
|
||||
if (tid == 0)
|
||||
{
|
||||
val = row[max(x - 1, 1)];
|
||||
sh_row[0] = val.x;
|
||||
sh_row[(NTHREADS + 2)] = val.y;
|
||||
sh_row[2 * (NTHREADS + 2)] = val.z;
|
||||
}
|
||||
|
||||
if (tid == gSizeX - 1)
|
||||
{
|
||||
val = row[min(x + 1, width - 2)];
|
||||
sh_row[gSizeX + 1] = val.x;
|
||||
sh_row[gSizeX + 1 + (NTHREADS + 2)] = val.y;
|
||||
sh_row[gSizeX + 1 + 2 * (NTHREADS + 2)] = val.z;
|
||||
}
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (x < width)
|
||||
{
|
||||
float4 a = (float4) (sh_row[tid], sh_row[tid + (NTHREADS + 2)],
|
||||
sh_row[tid + 2 * (NTHREADS + 2)], 0);
|
||||
float4 b = (float4) (sh_row[tid + 2], sh_row[tid + 2 + (NTHREADS + 2)],
|
||||
sh_row[tid + 2 + 2 * (NTHREADS + 2)], 0);
|
||||
|
||||
float4 dx;
|
||||
if (correct_gamma == 1)
|
||||
dx = sqrt(b) - sqrt(a);
|
||||
else
|
||||
dx = b - a;
|
||||
|
||||
float4 dy = (float4) 0.f;
|
||||
|
||||
if (gidY > 0 && gidY < height - 1)
|
||||
{
|
||||
a = convert_float4(img[(gidY - 1) * img_step + x].xyzw);
|
||||
b = convert_float4(img[(gidY + 1) * img_step + x].xyzw);
|
||||
|
||||
if (correct_gamma == 1)
|
||||
dy = sqrt(b) - sqrt(a);
|
||||
else
|
||||
dy = b - a;
|
||||
}
|
||||
|
||||
float4 mag = hypot(dx, dy);
|
||||
float best_dx = dx.x;
|
||||
float best_dy = dy.x;
|
||||
|
||||
float mag0 = mag.x;
|
||||
if (mag0 < mag.y)
|
||||
{
|
||||
best_dx = dx.y;
|
||||
best_dy = dy.y;
|
||||
mag0 = mag.y;
|
||||
}
|
||||
|
||||
if (mag0 < mag.z)
|
||||
{
|
||||
best_dx = dx.z;
|
||||
best_dy = dy.z;
|
||||
mag0 = mag.z;
|
||||
}
|
||||
|
||||
float ang = (atan2(best_dy, best_dx) + CV_PI_F) * angle_scale - 0.5f;
|
||||
int hidx = (int)floor(ang);
|
||||
ang -= hidx;
|
||||
hidx = (hidx + cnbins) % cnbins;
|
||||
|
||||
qangle[(gidY * qangle_step + x) << 1] = hidx;
|
||||
qangle[((gidY * qangle_step + x) << 1) + 1] = (hidx + 1) % cnbins;
|
||||
grad[(gidY * grad_quadstep + x) << 1] = mag0 * (1.f - ang);
|
||||
grad[((gidY * grad_quadstep + x) << 1) + 1] = mag0 * ang;
|
||||
}
|
||||
}
|
||||
|
||||
__kernel void compute_gradients_8UC1_kernel(
|
||||
const int height, const int width,
|
||||
const int img_step, const int grad_quadstep, const int qangle_step,
|
||||
__global const uchar * img, __global float * grad, __global QANGLE_TYPE * qangle,
|
||||
const float angle_scale, const char correct_gamma, const int cnbins)
|
||||
{
|
||||
const int x = get_global_id(0);
|
||||
const int tid = get_local_id(0);
|
||||
const int gSizeX = get_local_size(0);
|
||||
const int gidY = get_group_id(1);
|
||||
|
||||
__global const uchar* row = img + gidY * img_step;
|
||||
|
||||
__local float sh_row[NTHREADS + 2];
|
||||
|
||||
if (x < width)
|
||||
sh_row[tid + 1] = row[x];
|
||||
else
|
||||
sh_row[tid + 1] = row[width - 2];
|
||||
|
||||
if (tid == 0)
|
||||
sh_row[0] = row[max(x - 1, 1)];
|
||||
|
||||
if (tid == gSizeX - 1)
|
||||
sh_row[gSizeX + 1] = row[min(x + 1, width - 2)];
|
||||
|
||||
barrier(CLK_LOCAL_MEM_FENCE);
|
||||
if (x < width)
|
||||
{
|
||||
float dx;
|
||||
|
||||
if (correct_gamma == 1)
|
||||
dx = sqrt(sh_row[tid + 2]) - sqrt(sh_row[tid]);
|
||||
else
|
||||
dx = sh_row[tid + 2] - sh_row[tid];
|
||||
|
||||
float dy = 0.f;
|
||||
if (gidY > 0 && gidY < height - 1)
|
||||
{
|
||||
float a = (float) img[ (gidY + 1) * img_step + x ];
|
||||
float b = (float) img[ (gidY - 1) * img_step + x ];
|
||||
if (correct_gamma == 1)
|
||||
dy = sqrt(a) - sqrt(b);
|
||||
else
|
||||
dy = a - b;
|
||||
}
|
||||
float mag = hypot(dx, dy);
|
||||
|
||||
float ang = (atan2(dy, dx) + CV_PI_F) * angle_scale - 0.5f;
|
||||
int hidx = (int)floor(ang);
|
||||
ang -= hidx;
|
||||
hidx = (hidx + cnbins) % cnbins;
|
||||
|
||||
qangle[ (gidY * qangle_step + x) << 1 ] = hidx;
|
||||
qangle[ ((gidY * qangle_step + x) << 1) + 1 ] = (hidx + 1) % cnbins;
|
||||
grad[ (gidY * grad_quadstep + x) << 1 ] = mag * (1.f - ang);
|
||||
grad[ ((gidY * grad_quadstep + x) << 1) + 1 ] = mag * ang;
|
||||
}
|
||||
}
|
||||
53
Lib/opencv/sources/modules/objdetect/src/precomp.hpp
Normal file
53
Lib/opencv/sources/modules/objdetect/src/precomp.hpp
Normal file
@@ -0,0 +1,53 @@
|
||||
/*M///////////////////////////////////////////////////////////////////////////////////////
|
||||
//
|
||||
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
|
||||
//
|
||||
// By downloading, copying, installing or using the software you agree to this license.
|
||||
// If you do not agree to this license, do not download, install,
|
||||
// copy or use the software.
|
||||
//
|
||||
//
|
||||
// License Agreement
|
||||
// For Open Source Computer Vision Library
|
||||
//
|
||||
// Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
|
||||
// Copyright (C) 2009, Willow Garage Inc., all rights reserved.
|
||||
// Third party copyrights are property of their respective owners.
|
||||
//
|
||||
// Redistribution and use in source and binary forms, with or without modification,
|
||||
// are permitted provided that the following conditions are met:
|
||||
//
|
||||
// * Redistribution's of source code must retain the above copyright notice,
|
||||
// this list of conditions and the following disclaimer.
|
||||
//
|
||||
// * Redistribution's in binary form must reproduce the above copyright notice,
|
||||
// this list of conditions and the following disclaimer in the documentation
|
||||
// and/or other materials provided with the distribution.
|
||||
//
|
||||
// * The name of the copyright holders may not be used to endorse or promote products
|
||||
// derived from this software without specific prior written permission.
|
||||
//
|
||||
// This software is provided by the copyright holders and contributors "as is" and
|
||||
// any express or implied warranties, including, but not limited to, the implied
|
||||
// warranties of merchantability and fitness for a particular purpose are disclaimed.
|
||||
// In no event shall the Intel Corporation or contributors be liable for any direct,
|
||||
// indirect, incidental, special, exemplary, or consequential damages
|
||||
// (including, but not limited to, procurement of substitute goods or services;
|
||||
// loss of use, data, or profits; or business interruption) however caused
|
||||
// and on any theory of liability, whether in contract, strict liability,
|
||||
// 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.
|
||||
//
|
||||
//M*/
|
||||
|
||||
#ifndef __OPENCV_PRECOMP_H__
|
||||
#define __OPENCV_PRECOMP_H__
|
||||
|
||||
#include "opencv2/objdetect.hpp"
|
||||
#include "opencv2/imgproc.hpp"
|
||||
|
||||
#include "opencv2/core/utility.hpp"
|
||||
#include "opencv2/core/ocl.hpp"
|
||||
#include "opencv2/core/private.hpp"
|
||||
|
||||
#endif
|
||||
1232
Lib/opencv/sources/modules/objdetect/src/qrcode.cpp
Normal file
1232
Lib/opencv/sources/modules/objdetect/src/qrcode.cpp
Normal file
File diff suppressed because it is too large
Load Diff
Reference in New Issue
Block a user