Skip to content

O
opencv
Commit 1b7ad93d authored by marina.kolpakova's avatar marina.kolpakova
Browse files
Options

LBP features integrated in CascadeClassifier_GPU

parent 2dc93574
Show whitespace changes
Inline Side-by-side
Showing with 395 additions and 382 deletions
modules/gpu/include/opencv2/gpu/gpu.hpp
View file @ 1b7ad93d
...@@ -1397,7 +1397,7 @@ public: ...@@ -1397,7 +1397,7 @@ public:
}; };
////////////////////////////////// CascadeClassifier_GPU ////////////////////////////////////////// ////////////////////////////////// CascadeClassifier_GPU //////////////////////////////////////////
// The cascade classifier class for object detection. // The cascade classifier class for object detection: supports old haar and new lbp xlm formats and nvbin for haar cascades olny.
class CV_EXPORTS CascadeClassifier_GPU class CV_EXPORTS CascadeClassifier_GPU
{ {
public: public:
...@@ -1410,36 +1410,22 @@ public: ...@@ -1410,36 +1410,22 @@ public:
void release(); void release();
/* returns number of detected objects */ /* returns number of detected objects */
int detectMultiScale(const GpuMat& image, GpuMat& objectsBuf, double scaleFactor=1.2, int minNeighbors=4, Size minSize=Size()); int detectMultiScale(const GpuMat& image, GpuMat& objectsBuf, double scaleFactor = 1.2, int minNeighbors = 4, Size minSize = Size());
bool findLargestObject; bool findLargestObject;
bool visualizeInPlace; bool visualizeInPlace;
Size getClassifierSize() const; Size getClassifierSize() const;
private:
private:
struct CascadeClassifierImpl; struct CascadeClassifierImpl;
CascadeClassifierImpl* impl; CascadeClassifierImpl* impl;
}; struct HaarCascade;
struct LbpCascade;
friend class CascadeClassifier_GPU_LBP;
// The cascade classifier class for object detection.
class CV_EXPORTS CascadeClassifier_GPU_LBP
{
public: public:
CascadeClassifier_GPU_LBP(cv::Size detectionFrameSize = cv::Size()); int detectMultiScale(const GpuMat& image, GpuMat& objectsBuf, Size maxObjectSize, Size minSize = Size(), double scaleFactor = 1.1, int minNeighbors = 4);
~CascadeClassifier_GPU_LBP();
bool empty() const;
bool load(const std::string& filename);
void release();
int detectMultiScale(const GpuMat& image, GpuMat& objectsBuf, double scaleFactor = 1.1, int minNeighbors = 4,
cv::Size maxObjectSize = cv::Size()/*, Size minSize = Size()*/);
Size getClassifierSize() const;
private:
struct CascadeClassifierImpl;
CascadeClassifierImpl* impl;
}; };
////////////////////////////////// SURF ////////////////////////////////////////// ////////////////////////////////// SURF //////////////////////////////////////////
......
modules/gpu/perf/perf_objdetect.cpp
View file @ 1b7ad93d
...@@ -70,7 +70,7 @@ GPU_PERF_TEST_1(LBPClassifier, cv::gpu::DeviceInfo) ...@@ -70,7 +70,7 @@ GPU_PERF_TEST_1(LBPClassifier, cv::gpu::DeviceInfo)
cv::gpu::GpuMat img(img_host); cv::gpu::GpuMat img(img_host);
cv::gpu::GpuMat gpu_rects; cv::gpu::GpuMat gpu_rects;
cv::gpu::CascadeClassifier_GPU_LBP cascade(img.size()); cv::gpu::CascadeClassifier_GPU cascade;
ASSERT_TRUE(cascade.load(perf::TestBase::getDataPath("gpu/lbpcascade/lbpcascade_frontalface.xml"))); ASSERT_TRUE(cascade.load(perf::TestBase::getDataPath("gpu/lbpcascade/lbpcascade_frontalface.xml")));
cascade.detectMultiScale(img, gpu_rects); cascade.detectMultiScale(img, gpu_rects);
......
modules/gpu/src/cascadeclassifier.cpp
View file @ 1b7ad93d
...@@ -49,30 +49,238 @@ using namespace cv::gpu; ...@@ -49,30 +49,238 @@ using namespace cv::gpu;
using namespace std; using namespace std;
#if !defined (HAVE_CUDA) #if !defined (HAVE_CUDA)
// ============ old fashioned haar cascade ==============================================//
cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU() { throw_nogpu(); } cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU() { throw_nogpu(); }
cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU(const string&) { throw_nogpu(); } cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU(const string&) { throw_nogpu(); }
cv::gpu::CascadeClassifier_GPU::~CascadeClassifier_GPU() { throw_nogpu(); } cv::gpu::CascadeClassifier_GPU::~CascadeClassifier_GPU() { throw_nogpu(); }
bool cv::gpu::CascadeClassifier_GPU::empty() const { throw_nogpu(); return true; } bool cv::gpu::CascadeClassifier_GPU::empty() const { throw_nogpu(); return true; }
bool cv::gpu::CascadeClassifier_GPU::load(const string&) { throw_nogpu(); return true; } bool cv::gpu::CascadeClassifier_GPU::load(const string&) { throw_nogpu(); return true; }
Size cv::gpu::CascadeClassifier_GPU::getClassifierSize() const { throw_nogpu(); return Size(); } Size cv::gpu::CascadeClassifier_GPU::getClassifierSize() const { throw_nogpu(); return Size();}
void cv::gpu::CascadeClassifier_GPU::release() { throw_nogpu(); }
int cv::gpu::CascadeClassifier_GPU::detectMultiScale( const GpuMat&, GpuMat&, double, int, Size) {throw_nogpu(); return -1;}
int cv::gpu::CascadeClassifier_GPU::detectMultiScale( const GpuMat&, GpuMat&, Size, Size, double, int) {throw_nogpu(); return -1;}
int cv::gpu::CascadeClassifier_GPU::detectMultiScale( const GpuMat& , GpuMat& , double , int , Size) { throw_nogpu(); return 0; } #else
// ============ LBP cascade ==============================================// struct cv::gpu::CascadeClassifier_GPU::CascadeClassifierImpl
cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifier_GPU_LBP(cv::Size /*frameSize*/){ throw_nogpu(); } {
cv::gpu::CascadeClassifier_GPU_LBP::~CascadeClassifier_GPU_LBP() { throw_nogpu(); } public:
CascadeClassifierImpl(){}
virtual ~CascadeClassifierImpl(){}
bool cv::gpu::CascadeClassifier_GPU_LBP::empty() const { throw_nogpu(); return true; } virtual unsigned int process(const GpuMat& src, GpuMat& objects, float scaleStep, int minNeighbors,
bool cv::gpu::CascadeClassifier_GPU_LBP::load(const string&) { throw_nogpu(); return true; } bool findLargestObject, bool visualizeInPlace, cv::Size ncvMinSize, cv::Size maxObjectSize) = 0;
Size cv::gpu::CascadeClassifier_GPU_LBP::getClassifierSize() const { throw_nogpu(); return Size(); }
void cv::gpu::CascadeClassifier_GPU_LBP::allocateBuffers(cv::Size /*frame*/) { throw_nogpu();}
int cv::gpu::CascadeClassifier_GPU_LBP::detectMultiScale(const cv::gpu::GpuMat& /*image*/, cv::gpu::GpuMat& /*objectsBuf*/, virtual cv::Size getClassifierCvSize() const = 0;
double /*scaleFactor*/, int /*minNeighbors*/, cv::Size /*maxObjectSize*/){ throw_nogpu(); return 0;} virtual bool read(const string& classifierAsXml) = 0;
};
#else struct cv::gpu::CascadeClassifier_GPU::HaarCascade : cv::gpu::CascadeClassifier_GPU::CascadeClassifierImpl
{
public:
HaarCascade() : lastAllocatedFrameSize(-1, -1)
{
ncvSetDebugOutputHandler(NCVDebugOutputHandler);
}
bool read(const string& filename)
{
ncvSafeCall( load(filename) );
return true;
}
NCVStatus process(const GpuMat& src, GpuMat& objects, float scaleStep, int minNeighbors,
bool findLargestObject, bool visualizeInPlace, cv::Size ncvMinSize,
/*out*/unsigned int& numDetections)
{
calculateMemReqsAndAllocate(src.size());
NCVMemPtr src_beg;
src_beg.ptr = (void*)src.ptr<Ncv8u>();
src_beg.memtype = NCVMemoryTypeDevice;
NCVMemSegment src_seg;
src_seg.begin = src_beg;
src_seg.size = src.step * src.rows;
NCVMatrixReuse<Ncv8u> d_src(src_seg, static_cast<int>(devProp.textureAlignment), src.cols, src.rows, static_cast<int>(src.step), true);
ncvAssertReturn(d_src.isMemReused(), NCV_ALLOCATOR_BAD_REUSE);
CV_Assert(objects.rows == 1);
NCVMemPtr objects_beg;
objects_beg.ptr = (void*)objects.ptr<NcvRect32u>();
objects_beg.memtype = NCVMemoryTypeDevice;
NCVMemSegment objects_seg;
objects_seg.begin = objects_beg;
objects_seg.size = objects.step * objects.rows;
NCVVectorReuse<NcvRect32u> d_rects(objects_seg, objects.cols);
ncvAssertReturn(d_rects.isMemReused(), NCV_ALLOCATOR_BAD_REUSE);
NcvSize32u roi;
roi.width = d_src.width();
roi.height = d_src.height();
NcvSize32u winMinSize(ncvMinSize.width, ncvMinSize.height);
Ncv32u flags = 0;
flags |= findLargestObject? NCVPipeObjDet_FindLargestObject : 0;
flags |= visualizeInPlace ? NCVPipeObjDet_VisualizeInPlace : 0;
ncvStat = ncvDetectObjectsMultiScale_device(
d_src, roi, d_rects, numDetections, haar, *h_haarStages,
*d_haarStages, *d_haarNodes, *d_haarFeatures,
winMinSize,
minNeighbors,
scaleStep, 1,
flags,
*gpuAllocator, *cpuAllocator, devProp, 0);
ncvAssertReturnNcvStat(ncvStat);
ncvAssertCUDAReturn(cudaStreamSynchronize(0), NCV_CUDA_ERROR);
return NCV_SUCCESS;
}
unsigned int process(const GpuMat& image, GpuMat& objectsBuf, float scaleFactor, int minNeighbors,
bool findLargestObject, bool visualizeInPlace, cv::Size minSize, cv::Size maxObjectSize)
{
CV_Assert( scaleFactor > 1 && image.depth() == CV_8U);
const int defaultObjSearchNum = 100;
if (objectsBuf.empty())
{
objectsBuf.create(1, defaultObjSearchNum, DataType<Rect>::type);
}
cv::Size ncvMinSize = this->getClassifierCvSize();
if (ncvMinSize.width < (unsigned)minSize.width && ncvMinSize.height < (unsigned)minSize.height)
{
ncvMinSize.width = minSize.width;
ncvMinSize.height = minSize.height;
}
unsigned int numDetections;
ncvSafeCall(this->process(image, objectsBuf, (float)scaleFactor, minNeighbors, findLargestObject, visualizeInPlace, ncvMinSize, numDetections));
return numDetections;
}
cv::Size getClassifierCvSize() const { return cv::Size(haar.ClassifierSize.width, haar.ClassifierSize.height); }
private:
static void NCVDebugOutputHandler(const std::string &msg) { CV_Error(CV_GpuApiCallError, msg.c_str()); }
NCVStatus load(const string& classifierFile)
{
int devId = cv::gpu::getDevice();
ncvAssertCUDAReturn(cudaGetDeviceProperties(&devProp, devId), NCV_CUDA_ERROR);
// Load the classifier from file (assuming its size is about 1 mb) using a simple allocator
gpuCascadeAllocator = new NCVMemNativeAllocator(NCVMemoryTypeDevice, static_cast<int>(devProp.textureAlignment));
cpuCascadeAllocator = new NCVMemNativeAllocator(NCVMemoryTypeHostPinned, static_cast<int>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuCascadeAllocator->isInitialized(), "Error creating cascade GPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(cpuCascadeAllocator->isInitialized(), "Error creating cascade CPU allocator", NCV_CUDA_ERROR);
Ncv32u haarNumStages, haarNumNodes, haarNumFeatures;
ncvStat = ncvHaarGetClassifierSize(classifierFile, haarNumStages, haarNumNodes, haarNumFeatures);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error reading classifier size (check the file)", NCV_FILE_ERROR);
h_haarStages = new NCVVectorAlloc<HaarStage64>(*cpuCascadeAllocator, haarNumStages);
h_haarNodes = new NCVVectorAlloc<HaarClassifierNode128>(*cpuCascadeAllocator, haarNumNodes);
h_haarFeatures = new NCVVectorAlloc<HaarFeature64>(*cpuCascadeAllocator, haarNumFeatures);
ncvAssertPrintReturn(h_haarStages->isMemAllocated(), "Error in cascade CPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(h_haarNodes->isMemAllocated(), "Error in cascade CPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(h_haarFeatures->isMemAllocated(), "Error in cascade CPU allocator", NCV_CUDA_ERROR);
ncvStat = ncvHaarLoadFromFile_host(classifierFile, haar, *h_haarStages, *h_haarNodes, *h_haarFeatures);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error loading classifier", NCV_FILE_ERROR);
d_haarStages = new NCVVectorAlloc<HaarStage64>(*gpuCascadeAllocator, haarNumStages);
d_haarNodes = new NCVVectorAlloc<HaarClassifierNode128>(*gpuCascadeAllocator, haarNumNodes);
d_haarFeatures = new NCVVectorAlloc<HaarFeature64>(*gpuCascadeAllocator, haarNumFeatures);
ncvAssertPrintReturn(d_haarStages->isMemAllocated(), "Error in cascade GPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(d_haarNodes->isMemAllocated(), "Error in cascade GPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(d_haarFeatures->isMemAllocated(), "Error in cascade GPU allocator", NCV_CUDA_ERROR);
ncvStat = h_haarStages->copySolid(*d_haarStages, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", NCV_CUDA_ERROR);
ncvStat = h_haarNodes->copySolid(*d_haarNodes, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", NCV_CUDA_ERROR);
ncvStat = h_haarFeatures->copySolid(*d_haarFeatures, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", NCV_CUDA_ERROR);
return NCV_SUCCESS;
}
NCVStatus calculateMemReqsAndAllocate(const Size& frameSize)
{
if (lastAllocatedFrameSize == frameSize)
{
return NCV_SUCCESS;
}
// Calculate memory requirements and create real allocators
NCVMemStackAllocator gpuCounter(static_cast<int>(devProp.textureAlignment));
NCVMemStackAllocator cpuCounter(static_cast<int>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuCounter.isInitialized(), "Error creating GPU memory counter", NCV_CUDA_ERROR);
ncvAssertPrintReturn(cpuCounter.isInitialized(), "Error creating CPU memory counter", NCV_CUDA_ERROR);
NCVMatrixAlloc<Ncv8u> d_src(gpuCounter, frameSize.width, frameSize.height);
NCVMatrixAlloc<Ncv8u> h_src(cpuCounter, frameSize.width, frameSize.height);
ncvAssertReturn(d_src.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
ncvAssertReturn(h_src.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
NCVVectorAlloc<NcvRect32u> d_rects(gpuCounter, 100);
ncvAssertReturn(d_rects.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
NcvSize32u roi;
roi.width = d_src.width();
roi.height = d_src.height();
Ncv32u numDetections;
ncvStat = ncvDetectObjectsMultiScale_device(d_src, roi, d_rects, numDetections, haar, *h_haarStages,
*d_haarStages, *d_haarNodes, *d_haarFeatures, haar.ClassifierSize, 4, 1.2f, 1, 0, gpuCounter, cpuCounter, devProp, 0);
ncvAssertReturnNcvStat(ncvStat);
ncvAssertCUDAReturn(cudaStreamSynchronize(0), NCV_CUDA_ERROR);
gpuAllocator = new NCVMemStackAllocator(NCVMemoryTypeDevice, gpuCounter.maxSize(), static_cast<int>(devProp.textureAlignment));
cpuAllocator = new NCVMemStackAllocator(NCVMemoryTypeHostPinned, cpuCounter.maxSize(), static_cast<int>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuAllocator->isInitialized(), "Error creating GPU memory allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(cpuAllocator->isInitialized(), "Error creating CPU memory allocator", NCV_CUDA_ERROR);
return NCV_SUCCESS;
}
cudaDeviceProp devProp;
NCVStatus ncvStat;
Ptr<NCVMemNativeAllocator> gpuCascadeAllocator;
Ptr<NCVMemNativeAllocator> cpuCascadeAllocator;
Ptr<NCVVectorAlloc<HaarStage64> > h_haarStages;
Ptr<NCVVectorAlloc<HaarClassifierNode128> > h_haarNodes;
Ptr<NCVVectorAlloc<HaarFeature64> > h_haarFeatures;
HaarClassifierCascadeDescriptor haar;
Ptr<NCVVectorAlloc<HaarStage64> > d_haarStages;
Ptr<NCVVectorAlloc<HaarClassifierNode128> > d_haarNodes;
Ptr<NCVVectorAlloc<HaarFeature64> > d_haarFeatures;
Size lastAllocatedFrameSize;
Ptr<NCVMemStackAllocator> gpuAllocator;
Ptr<NCVMemStackAllocator> cpuAllocator;
virtual ~HaarCascade(){}
};
cv::Size operator -(const cv::Size& a, const cv::Size& b) cv::Size operator -(const cv::Size& a, const cv::Size& b)
{ {
...@@ -101,12 +309,17 @@ bool operator <=(const cv::Size& a, const cv::Size& b) ...@@ -101,12 +309,17 @@ bool operator <=(const cv::Size& a, const cv::Size& b)
struct PyrLavel struct PyrLavel
{ {
PyrLavel(int _order, float _scale, cv::Size frame, cv::Size window) : order(_order) PyrLavel(int _order, float _scale, cv::Size frame, cv::Size window, cv::Size minObjectSize)
{
do
{ {
order = _order;
scale = pow(_scale, order); scale = pow(_scale, order);
sFrame = frame / scale; sFrame = frame / scale;
workArea = sFrame - window + 1; workArea = sFrame - window + 1;
sWindow = window * scale; sWindow = window * scale;
_order++;
} while (sWindow <= minObjectSize);
} }
bool isFeasible(cv::Size maxObj) bool isFeasible(cv::Size maxObj)
...@@ -114,9 +327,9 @@ struct PyrLavel ...@@ -114,9 +327,9 @@ struct PyrLavel
return workArea.width > 0 && workArea.height > 0 && sWindow <= maxObj; return workArea.width > 0 && workArea.height > 0 && sWindow <= maxObj;
} }
PyrLavel next(float factor, cv::Size frame, cv::Size window) PyrLavel next(float factor, cv::Size frame, cv::Size window, cv::Size minObjectSize)
{ {
return PyrLavel(order + 1, factor, frame, window); return PyrLavel(order + 1, factor, frame, window, minObjectSize);
} }
int order; int order;
...@@ -152,7 +365,7 @@ namespace cv { namespace gpu { namespace device ...@@ -152,7 +365,7 @@ namespace cv { namespace gpu { namespace device
} }
}}} }}}
struct cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifierImpl struct cv::gpu::CascadeClassifier_GPU::LbpCascade : cv::gpu::CascadeClassifier_GPU::CascadeClassifierImpl
{ {
public: public:
struct Stage struct Stage
...@@ -162,55 +375,109 @@ public: ...@@ -162,55 +375,109 @@ public:
float threshold; float threshold;
}; };
bool read(const FileNode &root); LbpCascade(){}
void allocateBuffers(cv::Size frame = cv::Size()); virtual ~LbpCascade(){}
bool empty() const {return stage_mat.empty();}
int process(const GpuMat& image, GpuMat& objects, double scaleFactor, int groupThreshold, cv::Size maxObjectSize); virtual unsigned int process(const GpuMat& image, GpuMat& objects, float scaleFactor, int groupThreshold, bool findLargestObject,
bool visualizeInPlace, cv::Size minObjectSize, cv::Size maxObjectSize)
private: {
CV_Assert(scaleFactor > 1 && image.depth() == CV_8U);
enum stage { BOOST = 0 }; const int defaultObjSearchNum = 100;
enum feature { LBP = 0 }; const float grouping_eps = 0.2f;
static const stage stageType = BOOST; if( !objects.empty() && objects.depth() == CV_32S)
static const feature featureType = LBP; objects.reshape(4, 1);
else
objects.create(1 , image.cols >> 4, CV_32SC4);
cv::Size NxM; // used for debug
bool isStumps; // candidates.setTo(cv::Scalar::all(0));
int ncategories; // objects.setTo(cv::Scalar::all(0));
int subsetSize;
int nodeStep;
// gpu representation of classifier if (maxObjectSize == cv::Size())
GpuMat stage_mat; maxObjectSize = image.size();
GpuMat trees_mat;
GpuMat nodes_mat;
GpuMat leaves_mat;
GpuMat subsets_mat;
GpuMat features_mat;
GpuMat integral; allocateBuffers(image.size());
GpuMat integralBuffer;
GpuMat resuzeBuffer;
GpuMat candidates; unsigned int classified = 0;
static const int integralFactor = 4; GpuMat dclassified(1, 1, CV_32S);
}; cudaSafeCall( cudaMemcpy(dclassified.ptr(), &classified, sizeof(int), cudaMemcpyHostToDevice) );
void cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifierImpl::allocateBuffers(cv::Size frame) PyrLavel level(0, 1.0f, image.size(), NxM, minObjectSize);
{
if (frame == cv::Size())
return;
if (resuzeBuffer.empty() || frame.width > resuzeBuffer.cols || frame.height > resuzeBuffer.rows) while (level.isFeasible(maxObjectSize))
{ {
resuzeBuffer.create(frame, CV_8UC1); int acc = level.sFrame.width + 1;
float iniScale = level.scale;
integral.create(frame.height + 1, integralFactor * (frame.width + 1), CV_32SC1);
NcvSize32u roiSize; cv::Size area = level.workArea;
roiSize.width = frame.width; int step = 1 + (level.scale <= 2.f);
roiSize.height = frame.height;
int total = 0, prev = 0;
while (acc <= integralFactor * (image.cols + 1) && level.isFeasible(maxObjectSize))
{
// create sutable matrix headers
GpuMat src = resuzeBuffer(cv::Rect(0, 0, level.sFrame.width, level.sFrame.height));
GpuMat sint = integral(cv::Rect(prev, 0, level.sFrame.width + 1, level.sFrame.height + 1));
GpuMat buff = integralBuffer;
// generate integral for scale
gpu::resize(image, src, level.sFrame, 0, 0, CV_INTER_LINEAR);
gpu::integralBuffered(src, sint, buff);
// calculate job
int totalWidth = level.workArea.width / step;
total += totalWidth * (level.workArea.height / step);
// go to next pyramide level
level = level.next(scaleFactor, image.size(), NxM, minObjectSize);
area = level.workArea;
step = (1 + (level.scale <= 2.f));
prev = acc;
acc += level.sFrame.width + 1;
}
device::lbp::classifyPyramid(image.cols, image.rows, NxM.width - 1, NxM.height - 1, iniScale, scaleFactor, total, stage_mat, stage_mat.cols / sizeof(Stage), nodes_mat,
leaves_mat, subsets_mat, features_mat, subsetSize, candidates, dclassified.ptr<unsigned int>(), integral);
}
if (groupThreshold <= 0 || objects.empty())
return 0;
cudaSafeCall( cudaMemcpy(&classified, dclassified.ptr(), sizeof(int), cudaMemcpyDeviceToHost) );
device::lbp::connectedConmonents(candidates, classified, objects, groupThreshold, grouping_eps, dclassified.ptr<unsigned int>());
cudaSafeCall( cudaMemcpy(&classified, dclassified.ptr(), sizeof(int), cudaMemcpyDeviceToHost) );
cudaSafeCall( cudaDeviceSynchronize() );
return classified;
}
virtual cv::Size getClassifierCvSize() const { return NxM; }
bool read(const string& classifierAsXml)
{
FileStorage fs(classifierAsXml, FileStorage::READ);
return fs.isOpened() ? read(fs.getFirstTopLevelNode()) : false;
}
private:
void allocateBuffers(cv::Size frame)
{
if (frame == cv::Size())
return;
if (resuzeBuffer.empty() || frame.width > resuzeBuffer.cols || frame.height > resuzeBuffer.rows)
{
resuzeBuffer.create(frame, CV_8UC1);
integral.create(frame.height + 1, integralFactor * (frame.width + 1), CV_32SC1);
NcvSize32u roiSize;
roiSize.width = frame.width;
roiSize.height = frame.height;
cudaDeviceProp prop; cudaDeviceProp prop;
cudaSafeCall( cudaGetDeviceProperties(&prop, cv::gpu::getDevice()) ); cudaSafeCall( cudaGetDeviceProperties(&prop, cv::gpu::getDevice()) );
...@@ -221,11 +488,10 @@ void cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifierImpl::allocateBuffers( ...@@ -221,11 +488,10 @@ void cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifierImpl::allocateBuffers(
candidates.create(1 , frame.width >> 1, CV_32SC4); candidates.create(1 , frame.width >> 1, CV_32SC4);
} }
} }
// currently only stump based boost classifiers are supported bool read(const FileNode &root)
bool CascadeClassifier_GPU_LBP::CascadeClassifierImpl::read(const FileNode &root) {
{
const char *GPU_CC_STAGE_TYPE = "stageType"; const char *GPU_CC_STAGE_TYPE = "stageType";
const char *GPU_CC_FEATURE_TYPE = "featureType"; const char *GPU_CC_FEATURE_TYPE = "featureType";
const char *GPU_CC_BOOST = "BOOST"; const char *GPU_CC_BOOST = "BOOST";
...@@ -363,336 +629,97 @@ bool CascadeClassifier_GPU_LBP::CascadeClassifierImpl::read(const FileNode &root ...@@ -363,336 +629,97 @@ bool CascadeClassifier_GPU_LBP::CascadeClassifierImpl::read(const FileNode &root
features_mat.upload(cv::Mat(features).reshape(4,1)); features_mat.upload(cv::Mat(features).reshape(4,1));
return true; return true;
} }
int cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifierImpl::process(const GpuMat& image, GpuMat& objects, double scaleFactor, int groupThreshold, cv::Size maxObjectSize)
{
CV_Assert(!empty() && scaleFactor > 1 && image.depth() == CV_8U);
const int defaultObjSearchNum = 100;
const float grouping_eps = 0.2f;
if( !objects.empty() && objects.depth() == CV_32S)
objects.reshape(4, 1);
else
objects.create(1 , image.cols >> 4, CV_32SC4);
// used for debug
// candidates.setTo(cv::Scalar::all(0));
// objects.setTo(cv::Scalar::all(0));
if (maxObjectSize == cv::Size())
maxObjectSize = image.size();
allocateBuffers(image.size());
unsigned int classified = 0;
GpuMat dclassified(1, 1, CV_32S);
cudaSafeCall( cudaMemcpy(dclassified.ptr(), &classified, sizeof(int), cudaMemcpyHostToDevice) );
PyrLavel level(0, 1.0f, image.size(), NxM);
while (level.isFeasible(maxObjectSize))
{
int acc = level.sFrame.width + 1;
float iniScale = level.scale;
cv::Size area = level.workArea;
int step = 1 + (level.scale <= 2.f);
int total = 0, prev = 0;
while (acc <= integralFactor * (image.cols + 1) && level.isFeasible(maxObjectSize))
{
// create sutable matrix headers
GpuMat src = resuzeBuffer(cv::Rect(0, 0, level.sFrame.width, level.sFrame.height));
GpuMat sint = integral(cv::Rect(prev, 0, level.sFrame.width + 1, level.sFrame.height + 1));
GpuMat buff = integralBuffer;
// generate integral for scale
gpu::resize(image, src, level.sFrame, 0, 0, CV_INTER_LINEAR);
gpu::integralBuffered(src, sint, buff);
// calculate job enum stage { BOOST = 0 };
int totalWidth = level.workArea.width / step; enum feature { LBP = 1, HAAR = 2 };
// totalWidth = ((totalWidth + WARP_MASK) / WARP_SIZE) << WARP_LOG; static const stage stageType = BOOST;
static const feature featureType = LBP;
total += totalWidth * (level.workArea.height / step); cv::Size NxM;
bool isStumps;
int ncategories;
int subsetSize;
int nodeStep;
// go to next pyramide level // gpu representation of classifier
level = level.next(scaleFactor, image.size(), NxM); GpuMat stage_mat;
area = level.workArea; GpuMat trees_mat;
GpuMat nodes_mat;
GpuMat leaves_mat;
GpuMat subsets_mat;
GpuMat features_mat;
step = (1 + (level.scale <= 2.f)); GpuMat integral;
prev = acc; GpuMat integralBuffer;
acc += level.sFrame.width + 1; GpuMat resuzeBuffer;
}
device::lbp::classifyPyramid(image.cols, image.rows, NxM.width - 1, NxM.height - 1, iniScale, scaleFactor, total, stage_mat, stage_mat.cols / sizeof(Stage), nodes_mat, GpuMat candidates;
leaves_mat, subsets_mat, features_mat, subsetSize, candidates, dclassified.ptr<unsigned int>(), integral); static const int integralFactor = 4;
} };
if (groupThreshold <= 0 || objects.empty()) cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU()
return 0; : findLargestObject(false), visualizeInPlace(false), impl(0) {}
cudaSafeCall( cudaMemcpy(&classified, dclassified.ptr(), sizeof(int), cudaMemcpyDeviceToHost) ); cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU(const string& filename)
device::lbp::connectedConmonents(candidates, classified, objects, groupThreshold, grouping_eps, dclassified.ptr<unsigned int>()); : findLargestObject(false), visualizeInPlace(false), impl(0) { load(filename); }
// candidates.copyTo(objects); cv::gpu::CascadeClassifier_GPU::~CascadeClassifier_GPU() { release(); }
cudaSafeCall( cudaMemcpy(&classified, dclassified.ptr(), sizeof(int), cudaMemcpyDeviceToHost) );
cudaSafeCall( cudaDeviceSynchronize() );
return classified;
}
cv::gpu::CascadeClassifier_GPU_LBP::CascadeClassifier_GPU_LBP(cv::Size detectionFrameSize) : impl(new CascadeClassifierImpl()) { (*impl).allocateBuffers(detectionFrameSize); } void cv::gpu::CascadeClassifier_GPU::release() { if (impl) { delete impl; impl = 0; } }
cv::gpu::CascadeClassifier_GPU_LBP::~CascadeClassifier_GPU_LBP(){ delete impl; }
bool cv::gpu::CascadeClassifier_GPU::empty() const { return impl == 0; }
bool cv::gpu::CascadeClassifier_GPU_LBP::empty() const Size cv::gpu::CascadeClassifier_GPU::getClassifierSize() const
{ {
return (*impl).empty(); return this->empty() ? Size() : impl->getClassifierCvSize();
} }
bool cv::gpu::CascadeClassifier_GPU_LBP::load(const string& classifierAsXml) int cv::gpu::CascadeClassifier_GPU::detectMultiScale( const GpuMat& image, GpuMat& objectsBuf, double scaleFactor, int minNeighbors, Size minSize)
{ {
FileStorage fs(classifierAsXml, FileStorage::READ); CV_Assert( !this->empty());
return fs.isOpened() ? (*impl).read(fs.getFirstTopLevelNode()) : false; return impl->process(image, objectsBuf, (float)scaleFactor, minNeighbors, findLargestObject, visualizeInPlace, minSize, cv::Size());
} }
int cv::gpu::CascadeClassifier_GPU_LBP::detectMultiScale(const GpuMat& image, GpuMat& objects, double scaleFactor, int groupThreshold, cv::Size maxObjectSize) int cv::gpu::CascadeClassifier_GPU::detectMultiScale(const GpuMat& image, GpuMat& objectsBuf, Size maxObjectSize, Size minSize, double scaleFactor, int minNeighbors)
{ {
return (*impl).process(image, objects, scaleFactor, groupThreshold, maxObjectSize); CV_Assert( !this->empty());
return impl->process(image, objectsBuf, (float)scaleFactor, minNeighbors, findLargestObject, visualizeInPlace, minSize, maxObjectSize);
} }
// ============ old fashioned haar cascade ==============================================//
struct cv::gpu::CascadeClassifier_GPU::CascadeClassifierImpl
{
CascadeClassifierImpl(const string& filename) : lastAllocatedFrameSize(-1, -1)
{
ncvSetDebugOutputHandler(NCVDebugOutputHandler);
ncvSafeCall( load(filename) );
}
NCVStatus process(const GpuMat& src, GpuMat& objects, float scaleStep, int minNeighbors,
bool findLargestObject, bool visualizeInPlace, NcvSize32u ncvMinSize,
/*out*/unsigned int& numDetections)
{
calculateMemReqsAndAllocate(src.size());
NCVMemPtr src_beg;
src_beg.ptr = (void*)src.ptr<Ncv8u>();
src_beg.memtype = NCVMemoryTypeDevice;
NCVMemSegment src_seg;
src_seg.begin = src_beg;
src_seg.size = src.step * src.rows;
NCVMatrixReuse<Ncv8u> d_src(src_seg, static_cast<int>(devProp.textureAlignment), src.cols, src.rows, static_cast<int>(src.step), true);
ncvAssertReturn(d_src.isMemReused(), NCV_ALLOCATOR_BAD_REUSE);
CV_Assert(objects.rows == 1);
NCVMemPtr objects_beg;
objects_beg.ptr = (void*)objects.ptr<NcvRect32u>();
objects_beg.memtype = NCVMemoryTypeDevice;
NCVMemSegment objects_seg;
objects_seg.begin = objects_beg;
objects_seg.size = objects.step * objects.rows;
NCVVectorReuse<NcvRect32u> d_rects(objects_seg, objects.cols);
ncvAssertReturn(d_rects.isMemReused(), NCV_ALLOCATOR_BAD_REUSE);
NcvSize32u roi;
roi.width = d_src.width();
roi.height = d_src.height();
Ncv32u flags = 0;
flags |= findLargestObject? NCVPipeObjDet_FindLargestObject : 0;
flags |= visualizeInPlace ? NCVPipeObjDet_VisualizeInPlace : 0;
ncvStat = ncvDetectObjectsMultiScale_device(
d_src, roi, d_rects, numDetections, haar, *h_haarStages,
*d_haarStages, *d_haarNodes, *d_haarFeatures,
ncvMinSize,
minNeighbors,
scaleStep, 1,
flags,
*gpuAllocator, *cpuAllocator, devProp, 0);
ncvAssertReturnNcvStat(ncvStat);
ncvAssertCUDAReturn(cudaStreamSynchronize(0), NCV_CUDA_ERROR);
return NCV_SUCCESS;
}
NcvSize32u getClassifierSize() const { return haar.ClassifierSize; }
cv::Size getClassifierCvSize() const { return cv::Size(haar.ClassifierSize.width, haar.ClassifierSize.height); }
private:
static void NCVDebugOutputHandler(const std::string &msg) { CV_Error(CV_GpuApiCallError, msg.c_str()); }
NCVStatus load(const string& classifierFile)
{
int devId = cv::gpu::getDevice();
ncvAssertCUDAReturn(cudaGetDeviceProperties(&devProp, devId), NCV_CUDA_ERROR);
// Load the classifier from file (assuming its size is about 1 mb) using a simple allocator
gpuCascadeAllocator = new NCVMemNativeAllocator(NCVMemoryTypeDevice, static_cast<int>(devProp.textureAlignment));
cpuCascadeAllocator = new NCVMemNativeAllocator(NCVMemoryTypeHostPinned, static_cast<int>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuCascadeAllocator->isInitialized(), "Error creating cascade GPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(cpuCascadeAllocator->isInitialized(), "Error creating cascade CPU allocator", NCV_CUDA_ERROR);
Ncv32u haarNumStages, haarNumNodes, haarNumFeatures;
ncvStat = ncvHaarGetClassifierSize(classifierFile, haarNumStages, haarNumNodes, haarNumFeatures);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error reading classifier size (check the file)", NCV_FILE_ERROR);
h_haarStages = new NCVVectorAlloc<HaarStage64>(*cpuCascadeAllocator, haarNumStages);
h_haarNodes = new NCVVectorAlloc<HaarClassifierNode128>(*cpuCascadeAllocator, haarNumNodes);
h_haarFeatures = new NCVVectorAlloc<HaarFeature64>(*cpuCascadeAllocator, haarNumFeatures);
ncvAssertPrintReturn(h_haarStages->isMemAllocated(), "Error in cascade CPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(h_haarNodes->isMemAllocated(), "Error in cascade CPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(h_haarFeatures->isMemAllocated(), "Error in cascade CPU allocator", NCV_CUDA_ERROR);
ncvStat = ncvHaarLoadFromFile_host(classifierFile, haar, *h_haarStages, *h_haarNodes, *h_haarFeatures);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error loading classifier", NCV_FILE_ERROR);
d_haarStages = new NCVVectorAlloc<HaarStage64>(*gpuCascadeAllocator, haarNumStages);
d_haarNodes = new NCVVectorAlloc<HaarClassifierNode128>(*gpuCascadeAllocator, haarNumNodes);
d_haarFeatures = new NCVVectorAlloc<HaarFeature64>(*gpuCascadeAllocator, haarNumFeatures);
ncvAssertPrintReturn(d_haarStages->isMemAllocated(), "Error in cascade GPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(d_haarNodes->isMemAllocated(), "Error in cascade GPU allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(d_haarFeatures->isMemAllocated(), "Error in cascade GPU allocator", NCV_CUDA_ERROR);
ncvStat = h_haarStages->copySolid(*d_haarStages, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", NCV_CUDA_ERROR);
ncvStat = h_haarNodes->copySolid(*d_haarNodes, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", NCV_CUDA_ERROR);
ncvStat = h_haarFeatures->copySolid(*d_haarFeatures, 0);
ncvAssertPrintReturn(ncvStat == NCV_SUCCESS, "Error copying cascade to GPU", NCV_CUDA_ERROR);
return NCV_SUCCESS;
}
NCVStatus calculateMemReqsAndAllocate(const Size& frameSize)
{
if (lastAllocatedFrameSize == frameSize)
{
return NCV_SUCCESS;
}
// Calculate memory requirements and create real allocators
NCVMemStackAllocator gpuCounter(static_cast<int>(devProp.textureAlignment));
NCVMemStackAllocator cpuCounter(static_cast<int>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuCounter.isInitialized(), "Error creating GPU memory counter", NCV_CUDA_ERROR);
ncvAssertPrintReturn(cpuCounter.isInitialized(), "Error creating CPU memory counter", NCV_CUDA_ERROR);
NCVMatrixAlloc<Ncv8u> d_src(gpuCounter, frameSize.width, frameSize.height);
NCVMatrixAlloc<Ncv8u> h_src(cpuCounter, frameSize.width, frameSize.height);
ncvAssertReturn(d_src.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
ncvAssertReturn(h_src.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
NCVVectorAlloc<NcvRect32u> d_rects(gpuCounter, 100);
ncvAssertReturn(d_rects.isMemAllocated(), NCV_ALLOCATOR_BAD_ALLOC);
NcvSize32u roi;
roi.width = d_src.width();
roi.height = d_src.height();
Ncv32u numDetections;
ncvStat = ncvDetectObjectsMultiScale_device(d_src, roi, d_rects, numDetections, haar, *h_haarStages,
*d_haarStages, *d_haarNodes, *d_haarFeatures, haar.ClassifierSize, 4, 1.2f, 1, 0, gpuCounter, cpuCounter, devProp, 0);
ncvAssertReturnNcvStat(ncvStat);
ncvAssertCUDAReturn(cudaStreamSynchronize(0), NCV_CUDA_ERROR);
gpuAllocator = new NCVMemStackAllocator(NCVMemoryTypeDevice, gpuCounter.maxSize(), static_cast<int>(devProp.textureAlignment));
cpuAllocator = new NCVMemStackAllocator(NCVMemoryTypeHostPinned, cpuCounter.maxSize(), static_cast<int>(devProp.textureAlignment));
ncvAssertPrintReturn(gpuAllocator->isInitialized(), "Error creating GPU memory allocator", NCV_CUDA_ERROR);
ncvAssertPrintReturn(cpuAllocator->isInitialized(), "Error creating CPU memory allocator", NCV_CUDA_ERROR);
return NCV_SUCCESS;
}
cudaDeviceProp devProp;
NCVStatus ncvStat;
Ptr<NCVMemNativeAllocator> gpuCascadeAllocator;
Ptr<NCVMemNativeAllocator> cpuCascadeAllocator;
Ptr<NCVVectorAlloc<HaarStage64> > h_haarStages;
Ptr<NCVVectorAlloc<HaarClassifierNode128> > h_haarNodes;
Ptr<NCVVectorAlloc<HaarFeature64> > h_haarFeatures;
HaarClassifierCascadeDescriptor haar;
Ptr<NCVVectorAlloc<HaarStage64> > d_haarStages;
Ptr<NCVVectorAlloc<HaarClassifierNode128> > d_haarNodes;
Ptr<NCVVectorAlloc<HaarFeature64> > d_haarFeatures;
Size lastAllocatedFrameSize;
Ptr<NCVMemStackAllocator> gpuAllocator;
Ptr<NCVMemStackAllocator> cpuAllocator;
};
cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU() : findLargestObject(false), visualizeInPlace(false), impl(0) {}
cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU(const string& filename) : findLargestObject(false), visualizeInPlace(false), impl(0) { load(filename); }
cv::gpu::CascadeClassifier_GPU::~CascadeClassifier_GPU() { release(); }
bool cv::gpu::CascadeClassifier_GPU::empty() const { return impl == 0; }
void cv::gpu::CascadeClassifier_GPU::release() { if (impl) { delete impl; impl = 0; } }
bool cv::gpu::CascadeClassifier_GPU::load(const string& filename) bool cv::gpu::CascadeClassifier_GPU::load(const string& filename)
{ {
release(); release();
impl = new CascadeClassifierImpl(filename);
return !this->empty();
}
std::string fext = filename.substr(filename.find_last_of(".") + 1);
std::transform(fext.begin(), fext.end(), fext.begin(), ::tolower);
Size cv::gpu::CascadeClassifier_GPU::getClassifierSize() const if (fext == "nvbin")
{
return this->empty() ? Size() : impl->getClassifierCvSize();
}
int cv::gpu::CascadeClassifier_GPU::detectMultiScale( const GpuMat& image, GpuMat& objectsBuf, double scaleFactor, int minNeighbors, Size minSize)
{
CV_Assert( scaleFactor > 1 && image.depth() == CV_8U);
CV_Assert( !this->empty());
const int defaultObjSearchNum = 100;
if (objectsBuf.empty())
{ {
objectsBuf.create(1, defaultObjSearchNum, DataType<Rect>::type); impl = new HaarCascade();
return impl->read(filename);
} }
NcvSize32u ncvMinSize = impl->getClassifierSize(); FileStorage fs(filename, FileStorage::READ);
if (ncvMinSize.width < (unsigned)minSize.width && ncvMinSize.height < (unsigned)minSize.height) if (!fs.isOpened())
{ {
ncvMinSize.width = minSize.width; impl = new HaarCascade();
ncvMinSize.height = minSize.height; return impl->read(filename);
} }
unsigned int numDetections; const char *GPU_CC_LBP = "LBP";
ncvSafeCall( impl->process(image, objectsBuf, (float)scaleFactor, minNeighbors, findLargestObject, visualizeInPlace, ncvMinSize, numDetections) ); string featureTypeStr = (string)fs.getFirstTopLevelNode()["featureType"];
if (featureTypeStr == GPU_CC_LBP)
impl = new LbpCascade();
else
impl = new HaarCascade();
return numDetections; impl->read(filename);
return !this->empty();
} }
//////////////////////////////////////////////////////////////////////////////////////////////////////
struct RectConvert struct RectConvert
{ {
...@@ -708,7 +735,6 @@ struct RectConvert ...@@ -708,7 +735,6 @@ struct RectConvert
} }
}; };
void groupRectangles(std::vector<NcvRect32u> &hypotheses, int groupThreshold, double eps, std::vector<Ncv32u> *weights) void groupRectangles(std::vector<NcvRect32u> &hypotheses, int groupThreshold, double eps, std::vector<Ncv32u> *weights)
{ {
vector<Rect> rects(hypotheses.size()); vector<Rect> rects(hypotheses.size());
......
modules/gpu/src/cuda/lbp.cu
View file @ 1b7ad93d
...@@ -290,6 +290,7 @@ namespace cv { namespace gpu { namespace device ...@@ -290,6 +290,7 @@ namespace cv { namespace gpu { namespace device
{ {
const int block = 128; const int block = 128;
int grid = divUp(workAmount, block); int grid = divUp(workAmount, block);
cudaFuncSetCacheConfig(lbp_cascade, cudaFuncCachePreferL1);
Cascade cascade((Stage*)mstages.ptr(), nstages, (ClNode*)mnodes.ptr(), mleaves.ptr(), msubsets.ptr(), (uchar4*)mfeatures.ptr(), subsetSize); Cascade cascade((Stage*)mstages.ptr(), nstages, (ClNode*)mnodes.ptr(), mleaves.ptr(), msubsets.ptr(), (uchar4*)mfeatures.ptr(), subsetSize);
lbp_cascade<<<grid, block>>>(cascade, frameW, frameH, windowW, windowH, initialScale, factor, workAmount, integral.ptr(), integral.step / sizeof(int), objects, classified); lbp_cascade<<<grid, block>>>(cascade, frameW, frameH, windowW, windowH, initialScale, factor, workAmount, integral.ptr(), integral.step / sizeof(int), objects, classified);
} }
......
modules/gpu/test/test_objdetect.cpp
View file @ 1b7ad93d
...@@ -302,7 +302,7 @@ PARAM_TEST_CASE(LBP_Read_classifier, cv::gpu::DeviceInfo, int) ...@@ -302,7 +302,7 @@ PARAM_TEST_CASE(LBP_Read_classifier, cv::gpu::DeviceInfo, int)
TEST_P(LBP_Read_classifier, Accuracy) TEST_P(LBP_Read_classifier, Accuracy)
{ {
cv::gpu::CascadeClassifier_GPU_LBP classifier; cv::gpu::CascadeClassifier_GPU classifier;
std::string classifierXmlPath = std::string(cvtest::TS::ptr()->get_data_path()) + "lbpcascade/lbpcascade_frontalface.xml"; std::string classifierXmlPath = std::string(cvtest::TS::ptr()->get_data_path()) + "lbpcascade/lbpcascade_frontalface.xml";
ASSERT_TRUE(classifier.load(classifierXmlPath)); ASSERT_TRUE(classifier.load(classifierXmlPath));
} }
...@@ -344,7 +344,7 @@ TEST_P(LBP_classify, Accuracy) ...@@ -344,7 +344,7 @@ TEST_P(LBP_classify, Accuracy)
for (; it != rects.end(); ++it) for (; it != rects.end(); ++it)
cv::rectangle(markedImage, *it, CV_RGB(0, 0, 255)); cv::rectangle(markedImage, *it, CV_RGB(0, 0, 255));
cv::gpu::CascadeClassifier_GPU_LBP gpuClassifier; cv::gpu::CascadeClassifier_GPU gpuClassifier;
ASSERT_TRUE(gpuClassifier.load(classifierXmlPath)); ASSERT_TRUE(gpuClassifier.load(classifierXmlPath));
cv::gpu::GpuMat gpu_rects; cv::gpu::GpuMat gpu_rects;
...@@ -359,8 +359,8 @@ TEST_P(LBP_classify, Accuracy) ...@@ -359,8 +359,8 @@ TEST_P(LBP_classify, Accuracy)
#if defined (LOG_CASCADE_STATISTIC) #if defined (LOG_CASCADE_STATISTIC)
std::cout << r.x << " " << r.y << " " << r.width << " " << r.height << std::endl; std::cout << r.x << " " << r.y << " " << r.width << " " << r.height << std::endl;
#endif
cv::rectangle(markedImage, r , CV_RGB(255, 0, 0)); cv::rectangle(markedImage, r , CV_RGB(255, 0, 0));
#endif
} }
#if defined (LOG_CASCADE_STATISTIC) #if defined (LOG_CASCADE_STATISTIC)
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment