opencv/modules/gpu/src/cascadeclassifier.cpp

/*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 GpuMaterials 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 bpied warranties, including, but not limited to, the bpied
// 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 <vector>
#include <iostream>

using namespace cv;
using namespace cv::gpu;
using namespace std;

#if !defined (HAVE_CUDA)

cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU()               { throw_nogpu(); }
cv::gpu::CascadeClassifier_GPU::CascadeClassifier_GPU(const string&)  { 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::load(const string&)              { throw_nogpu(); return true; }
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;}

#else

struct cv::gpu::CascadeClassifier_GPU::CascadeClassifierImpl
{
public:
    CascadeClassifierImpl(){}
    virtual ~CascadeClassifierImpl(){}

    virtual unsigned int process(const GpuMat& src, GpuMat& objects, float scaleStep, int minNeighbors,
                      bool findLargestObject, bool visualizeInPlace, cv::Size ncvMinSize, cv::Size maxObjectSize) = 0;

    virtual cv::Size getClassifierCvSize() const = 0;
    virtual bool read(const string& classifierAsXml) = 0;
};

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 < minSize.width && ncvMinSize.height < 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);

        lastAllocatedFrameSize = frameSize;
        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)
{
    return cv::Size(a.width - b.width, a.height - b.height);
}

cv::Size operator +(const cv::Size& a, const int& i)
{
    return cv::Size(a.width + i, a.height + i);
}

cv::Size operator *(const cv::Size& a, const float& f)
{
    return cv::Size(cvRound(a.width * f), cvRound(a.height * f));
}

cv::Size operator /(const cv::Size& a, const float& f)
{
    return cv::Size(cvRound(a.width / f), cvRound(a.height / f));
}

bool operator <=(const cv::Size& a, const cv::Size& b)
{
    return a.width <= b.width && a.height <= b.width;
}

struct PyrLavel
{
    PyrLavel(int _order, float _scale, cv::Size frame, cv::Size window, cv::Size minObjectSize)
    {
        do
        {
            order = _order;
            scale = pow(_scale, order);
            sFrame = frame / scale;
            workArea = sFrame - window + 1;
            sWindow = window * scale;
            _order++;
        } while (sWindow <= minObjectSize);
    }

    bool isFeasible(cv::Size maxObj)
    {
        return workArea.width > 0 && workArea.height > 0 && sWindow <= maxObj;
    }

    PyrLavel next(float factor, cv::Size frame, cv::Size window, cv::Size minObjectSize)
    {
        return PyrLavel(order + 1, factor, frame, window, minObjectSize);
    }

    int order;
    float scale;
    cv::Size sFrame;
    cv::Size workArea;
    cv::Size sWindow;
};

namespace cv { namespace gpu { namespace device
{
    namespace lbp
    {
        void classifyPyramid(int frameW,
                             int frameH,
                             int windowW,
                             int windowH,
                             float initalScale,
                             float factor,
                             int total,
                             const PtrStepSzb& mstages,
                             const int nstages,
                             const PtrStepSzi& mnodes,
                             const PtrStepSzf& mleaves,
                             const PtrStepSzi& msubsets,
                             const PtrStepSzb& mfeatures,
                             const int subsetSize,
                             PtrStepSz<int4> objects,
                             unsigned int* classified,
                             PtrStepSzi integral);

        void connectedConmonents(PtrStepSz<int4>  candidates, int ncandidates, PtrStepSz<int4> objects,int groupThreshold, float grouping_eps, unsigned int* nclasses);
    }
}}}

struct cv::gpu::CascadeClassifier_GPU::LbpCascade : cv::gpu::CascadeClassifier_GPU::CascadeClassifierImpl
{
public:
    struct Stage
    {
        int    first;
        int    ntrees;
        float  threshold;
    };

    LbpCascade(){}
    virtual ~LbpCascade(){}

    virtual unsigned int process(const GpuMat& image, GpuMat& objects, float scaleFactor, int groupThreshold, bool /*findLargestObject*/,
        bool /*visualizeInPlace*/, cv::Size minObjectSize, cv::Size maxObjectSize)
    {
        CV_Assert(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, minObjectSize);

        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
                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;
            cudaSafeCall( cudaGetDeviceProperties(&prop, cv::gpu::getDevice()) );

            Ncv32u bufSize;
            ncvSafeCall( nppiStIntegralGetSize_8u32u(roiSize, &bufSize, prop) );
            integralBuffer.create(1, bufSize, CV_8UC1);

            candidates.create(1 , frame.width >> 1, CV_32SC4);
        }
    }

    bool read(const FileNode &root)
    {
        const char *GPU_CC_STAGE_TYPE       = "stageType";
        const char *GPU_CC_FEATURE_TYPE     = "featureType";
        const char *GPU_CC_BOOST            = "BOOST";
        const char *GPU_CC_LBP              = "LBP";
        const char *GPU_CC_MAX_CAT_COUNT    = "maxCatCount";
        const char *GPU_CC_HEIGHT           = "height";
        const char *GPU_CC_WIDTH            = "width";
        const char *GPU_CC_STAGE_PARAMS     = "stageParams";
        const char *GPU_CC_MAX_DEPTH        = "maxDepth";
        const char *GPU_CC_FEATURE_PARAMS   = "featureParams";
        const char *GPU_CC_STAGES           = "stages";
        const char *GPU_CC_STAGE_THRESHOLD  = "stageThreshold";
        const float GPU_THRESHOLD_EPS       = 1e-5f;
        const char *GPU_CC_WEAK_CLASSIFIERS = "weakClassifiers";
        const char *GPU_CC_INTERNAL_NODES   = "internalNodes";
        const char *GPU_CC_LEAF_VALUES      = "leafValues";
        const char *GPU_CC_FEATURES         = "features";
        const char *GPU_CC_RECT             = "rect";

        std::string stageTypeStr = (string)root[GPU_CC_STAGE_TYPE];
        CV_Assert(stageTypeStr == GPU_CC_BOOST);

        string featureTypeStr = (string)root[GPU_CC_FEATURE_TYPE];
        CV_Assert(featureTypeStr == GPU_CC_LBP);

        NxM.width =  (int)root[GPU_CC_WIDTH];
        NxM.height = (int)root[GPU_CC_HEIGHT];
        CV_Assert( NxM.height > 0 && NxM.width > 0 );

        isStumps = ((int)(root[GPU_CC_STAGE_PARAMS][GPU_CC_MAX_DEPTH]) == 1) ? true : false;
        CV_Assert(isStumps);

        FileNode fn = root[GPU_CC_FEATURE_PARAMS];
        if (fn.empty())
            return false;

        ncategories = fn[GPU_CC_MAX_CAT_COUNT];

        subsetSize = (ncategories + 31) / 32;
        nodeStep = 3 + ( ncategories > 0 ? subsetSize : 1 );

        fn = root[GPU_CC_STAGES];
        if (fn.empty())
            return false;

        std::vector<Stage> stages;
        stages.reserve(fn.size());

        std::vector<int> cl_trees;
        std::vector<int> cl_nodes;
        std::vector<float> cl_leaves;
        std::vector<int> subsets;

        FileNodeIterator it = fn.begin(), it_end = fn.end();
        for (size_t si = 0; it != it_end; si++, ++it )
        {
            FileNode fns = *it;
            Stage st;
            st.threshold = (float)fns[GPU_CC_STAGE_THRESHOLD] - GPU_THRESHOLD_EPS;

            fns = fns[GPU_CC_WEAK_CLASSIFIERS];
            if (fns.empty())
                return false;

            st.ntrees = (int)fns.size();
            st.first = (int)cl_trees.size();

            stages.push_back(st);// (int, int, float)

            cl_trees.reserve(stages[si].first + stages[si].ntrees);

            // weak trees
            FileNodeIterator it1 = fns.begin(), it1_end = fns.end();
            for ( ; it1 != it1_end; ++it1 )
            {
                FileNode fnw = *it1;

                FileNode internalNodes = fnw[GPU_CC_INTERNAL_NODES];
                FileNode leafValues = fnw[GPU_CC_LEAF_VALUES];
                if ( internalNodes.empty() || leafValues.empty() )
                    return false;

                int nodeCount = (int)internalNodes.size()/nodeStep;
                cl_trees.push_back(nodeCount);

                cl_nodes.reserve((cl_nodes.size() + nodeCount) * 3);
                cl_leaves.reserve(cl_leaves.size() + leafValues.size());

                if( subsetSize > 0 )
                    subsets.reserve(subsets.size() + nodeCount * subsetSize);

                // nodes
                FileNodeIterator iIt = internalNodes.begin(), iEnd = internalNodes.end();

                for( ; iIt != iEnd; )
                {
                    cl_nodes.push_back((int)*(iIt++));
                    cl_nodes.push_back((int)*(iIt++));
                    cl_nodes.push_back((int)*(iIt++));

                    if( subsetSize > 0 )
                        for( int j = 0; j < subsetSize; j++, ++iIt )
                            subsets.push_back((int)*iIt);
                }

                // leaves
                iIt = leafValues.begin(), iEnd = leafValues.end();
                for( ; iIt != iEnd; ++iIt )
                    cl_leaves.push_back((float)*iIt);
            }
        }

        fn = root[GPU_CC_FEATURES];
        if( fn.empty() )
            return false;
        std::vector<uchar> features;
        features.reserve(fn.size() * 4);
        FileNodeIterator f_it = fn.begin(), f_end = fn.end();
        for (; f_it != f_end; ++f_it)
        {
            FileNode rect = (*f_it)[GPU_CC_RECT];
            FileNodeIterator r_it = rect.begin();
            features.push_back(saturate_cast<uchar>((int)*(r_it++)));
            features.push_back(saturate_cast<uchar>((int)*(r_it++)));
            features.push_back(saturate_cast<uchar>((int)*(r_it++)));
            features.push_back(saturate_cast<uchar>((int)*(r_it++)));
        }

        // copy data structures on gpu
        stage_mat.upload(cv::Mat(1, stages.size() * sizeof(Stage), CV_8UC1, (uchar*)&(stages[0]) ));
        trees_mat.upload(cv::Mat(cl_trees).reshape(1,1));
        nodes_mat.upload(cv::Mat(cl_nodes).reshape(1,1));
        leaves_mat.upload(cv::Mat(cl_leaves).reshape(1,1));
        subsets_mat.upload(cv::Mat(subsets).reshape(1,1));
        features_mat.upload(cv::Mat(features).reshape(4,1));

        return true;
    }

    enum stage { BOOST = 0 };
    enum feature { LBP = 1, HAAR = 2 };
    static const stage stageType = BOOST;
    static const feature featureType = LBP;

    cv::Size NxM;
    bool isStumps;
    int ncategories;
    int subsetSize;
    int nodeStep;

    // gpu representation of classifier
    GpuMat stage_mat;
    GpuMat trees_mat;
    GpuMat nodes_mat;
    GpuMat leaves_mat;
    GpuMat subsets_mat;
    GpuMat features_mat;

    GpuMat integral;
    GpuMat integralBuffer;
    GpuMat resuzeBuffer;

    GpuMat candidates;
    static const int integralFactor = 4;
};

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(); }

void cv::gpu::CascadeClassifier_GPU::release() { if (impl) { delete impl; impl = 0; } }

bool cv::gpu::CascadeClassifier_GPU::empty() const { return impl == 0; }

Size cv::gpu::CascadeClassifier_GPU::getClassifierSize() const
{
    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( !this->empty());
    return impl->process(image, objectsBuf, (float)scaleFactor, minNeighbors, findLargestObject, visualizeInPlace, minSize, cv::Size());
}

int cv::gpu::CascadeClassifier_GPU::detectMultiScale(const GpuMat& image, GpuMat& objectsBuf, Size maxObjectSize, Size minSize, double scaleFactor, int minNeighbors)
{
    CV_Assert( !this->empty());
    return impl->process(image, objectsBuf, (float)scaleFactor, minNeighbors, findLargestObject, visualizeInPlace, minSize, maxObjectSize);
}

bool cv::gpu::CascadeClassifier_GPU::load(const string& filename)
{
    release();

    std::string fext = filename.substr(filename.find_last_of(".") + 1);
    std::transform(fext.begin(), fext.end(), fext.begin(), ::tolower);

    if (fext == "nvbin")
    {
        impl = new HaarCascade();
        return impl->read(filename);
    }

    FileStorage fs(filename, FileStorage::READ);

    if (!fs.isOpened())
    {
        impl = new HaarCascade();
        return impl->read(filename);
    }

    const char *GPU_CC_LBP = "LBP";
    string featureTypeStr = (string)fs.getFirstTopLevelNode()["featureType"];
    if (featureTypeStr == GPU_CC_LBP)
        impl = new LbpCascade();
    else
        impl = new HaarCascade();

    impl->read(filename);
    return !this->empty();
}

//////////////////////////////////////////////////////////////////////////////////////////////////////

struct RectConvert
{
    Rect operator()(const NcvRect32u& nr) const { return Rect(nr.x, nr.y, nr.width, nr.height); }
    NcvRect32u operator()(const Rect& nr) const
    {
        NcvRect32u rect;
        rect.x = nr.x;
        rect.y = nr.y;
        rect.width = nr.width;
        rect.height = nr.height;
        return rect;
    }
};

void groupRectangles(std::vector<NcvRect32u> &hypotheses, int groupThreshold, double eps, std::vector<Ncv32u> *weights)
{
    vector<Rect> rects(hypotheses.size());
    std::transform(hypotheses.begin(), hypotheses.end(), rects.begin(), RectConvert());

    if (weights)
    {
        vector<int> weights_int;
        weights_int.assign(weights->begin(), weights->end());
        cv::groupRectangles(rects, weights_int, groupThreshold, eps);
    }
    else
    {
        cv::groupRectangles(rects, groupThreshold, eps);
    }
    std::transform(rects.begin(), rects.end(), hypotheses.begin(), RectConvert());
    hypotheses.resize(rects.size());
}

NCVStatus loadFromXML(const std::string &filename,
                      HaarClassifierCascadeDescriptor &haar,
                      std::vector<HaarStage64> &haarStages,
                      std::vector<HaarClassifierNode128> &haarClassifierNodes,
                      std::vector<HaarFeature64> &haarFeatures)
{
    NCVStatus ncvStat;

    haar.NumStages = 0;
    haar.NumClassifierRootNodes = 0;
    haar.NumClassifierTotalNodes = 0;
    haar.NumFeatures = 0;
    haar.ClassifierSize.width = 0;
    haar.ClassifierSize.height = 0;
    haar.bHasStumpsOnly = true;
    haar.bNeedsTiltedII = false;
    Ncv32u curMaxTreeDepth;

    std::vector<char> xmlFileCont;

    std::vector<HaarClassifierNode128> h_TmpClassifierNotRootNodes;
    haarStages.resize(0);
    haarClassifierNodes.resize(0);
    haarFeatures.resize(0);

    Ptr<CvHaarClassifierCascade> oldCascade = (CvHaarClassifierCascade*)cvLoad(filename.c_str(), 0, 0, 0);
    if (oldCascade.empty())
    {
        return NCV_HAAR_XML_LOADING_EXCEPTION;
    }

    haar.ClassifierSize.width = oldCascade->orig_window_size.width;
    haar.ClassifierSize.height = oldCascade->orig_window_size.height;

    int stagesCound = oldCascade->count;
    for(int s = 0; s < stagesCound; ++s) // by stages
    {
        HaarStage64 curStage;
        curStage.setStartClassifierRootNodeOffset(static_cast<Ncv32u>(haarClassifierNodes.size()));

        curStage.setStageThreshold(oldCascade->stage_classifier[s].threshold);

        int treesCount = oldCascade->stage_classifier[s].count;
        for(int t = 0; t < treesCount; ++t) // by trees
        {
            Ncv32u nodeId = 0;
            CvHaarClassifier* tree = &oldCascade->stage_classifier[s].classifier[t];

            int nodesCount = tree->count;
            for(int n = 0; n < nodesCount; ++n)  //by features
            {
                CvHaarFeature* feature = &tree->haar_feature[n];

                HaarClassifierNode128 curNode;
                curNode.setThreshold(tree->threshold[n]);

                NcvBool bIsLeftNodeLeaf = false;
                NcvBool bIsRightNodeLeaf = false;

                HaarClassifierNodeDescriptor32 nodeLeft;
                if ( tree->left[n] <= 0 )
                {
                    Ncv32f leftVal = tree->alpha[-tree->left[n]];
                    ncvStat = nodeLeft.create(leftVal);
                    ncvAssertReturn(ncvStat == NCV_SUCCESS, ncvStat);
                    bIsLeftNodeLeaf = true;
                }
                else
                {
                    Ncv32u leftNodeOffset = tree->left[n];
                    nodeLeft.create((Ncv32u)(h_TmpClassifierNotRootNodes.size() + leftNodeOffset - 1));
                    haar.bHasStumpsOnly = false;
                }
                curNode.setLeftNodeDesc(nodeLeft);

                HaarClassifierNodeDescriptor32 nodeRight;
                if ( tree->right[n] <= 0 )
                {
                    Ncv32f rightVal = tree->alpha[-tree->right[n]];
                    ncvStat = nodeRight.create(rightVal);
                    ncvAssertReturn(ncvStat == NCV_SUCCESS, ncvStat);
                    bIsRightNodeLeaf = true;
                }
                else
                {
                    Ncv32u rightNodeOffset = tree->right[n];
                    nodeRight.create((Ncv32u)(h_TmpClassifierNotRootNodes.size() + rightNodeOffset - 1));
                    haar.bHasStumpsOnly = false;
                }
                curNode.setRightNodeDesc(nodeRight);

                Ncv32u tiltedVal = feature->tilted;
                haar.bNeedsTiltedII = (tiltedVal != 0);

                Ncv32u featureId = 0;
                for(int l = 0; l < CV_HAAR_FEATURE_MAX; ++l) //by rects
                {
                    Ncv32u rectX = feature->rect[l].r.x;
                    Ncv32u rectY = feature->rect[l].r.y;
                    Ncv32u rectWidth = feature->rect[l].r.width;
                    Ncv32u rectHeight = feature->rect[l].r.height;

                    Ncv32f rectWeight = feature->rect[l].weight;

                    if (rectWeight == 0/* && rectX == 0 &&rectY == 0 && rectWidth == 0 && rectHeight == 0*/)
                        break;

                    HaarFeature64 curFeature;
                    ncvStat = curFeature.setRect(rectX, rectY, rectWidth, rectHeight, haar.ClassifierSize.width, haar.ClassifierSize.height);
                    curFeature.setWeight(rectWeight);
                    ncvAssertReturn(NCV_SUCCESS == ncvStat, ncvStat);
                    haarFeatures.push_back(curFeature);

                    featureId++;
                }

                HaarFeatureDescriptor32 tmpFeatureDesc;
                ncvStat = tmpFeatureDesc.create(haar.bNeedsTiltedII, bIsLeftNodeLeaf, bIsRightNodeLeaf,
                    featureId, static_cast<Ncv32u>(haarFeatures.size()) - featureId);
                ncvAssertReturn(NCV_SUCCESS == ncvStat, ncvStat);
                curNode.setFeatureDesc(tmpFeatureDesc);

                if (!nodeId)
                {
                    //root node
                    haarClassifierNodes.push_back(curNode);
                    curMaxTreeDepth = 1;
                }
                else
                {
                    //other node
                    h_TmpClassifierNotRootNodes.push_back(curNode);
                    curMaxTreeDepth++;
                }

                nodeId++;
            }
        }

        curStage.setNumClassifierRootNodes(treesCount);
        haarStages.push_back(curStage);
    }

    //fill in cascade stats
    haar.NumStages = static_cast<Ncv32u>(haarStages.size());
    haar.NumClassifierRootNodes = static_cast<Ncv32u>(haarClassifierNodes.size());
    haar.NumClassifierTotalNodes = static_cast<Ncv32u>(haar.NumClassifierRootNodes + h_TmpClassifierNotRootNodes.size());
    haar.NumFeatures = static_cast<Ncv32u>(haarFeatures.size());

    //merge root and leaf nodes in one classifiers array
    Ncv32u offsetRoot = static_cast<Ncv32u>(haarClassifierNodes.size());
    for (Ncv32u i=0; i<haarClassifierNodes.size(); i++)
    {
        HaarFeatureDescriptor32 featureDesc = haarClassifierNodes[i].getFeatureDesc();

        HaarClassifierNodeDescriptor32 nodeLeft = haarClassifierNodes[i].getLeftNodeDesc();
        if (!featureDesc.isLeftNodeLeaf())
        {
            Ncv32u newOffset = nodeLeft.getNextNodeOffset() + offsetRoot;
            nodeLeft.create(newOffset);
        }
        haarClassifierNodes[i].setLeftNodeDesc(nodeLeft);

        HaarClassifierNodeDescriptor32 nodeRight = haarClassifierNodes[i].getRightNodeDesc();
        if (!featureDesc.isRightNodeLeaf())
        {
            Ncv32u newOffset = nodeRight.getNextNodeOffset() + offsetRoot;
            nodeRight.create(newOffset);
        }
        haarClassifierNodes[i].setRightNodeDesc(nodeRight);
    }

    for (Ncv32u i=0; i<h_TmpClassifierNotRootNodes.size(); i++)
    {
        HaarFeatureDescriptor32 featureDesc = h_TmpClassifierNotRootNodes[i].getFeatureDesc();

        HaarClassifierNodeDescriptor32 nodeLeft = h_TmpClassifierNotRootNodes[i].getLeftNodeDesc();
        if (!featureDesc.isLeftNodeLeaf())
        {
            Ncv32u newOffset = nodeLeft.getNextNodeOffset() + offsetRoot;
            nodeLeft.create(newOffset);
        }
        h_TmpClassifierNotRootNodes[i].setLeftNodeDesc(nodeLeft);

        HaarClassifierNodeDescriptor32 nodeRight = h_TmpClassifierNotRootNodes[i].getRightNodeDesc();
        if (!featureDesc.isRightNodeLeaf())
        {
            Ncv32u newOffset = nodeRight.getNextNodeOffset() + offsetRoot;
            nodeRight.create(newOffset);
        }
        h_TmpClassifierNotRootNodes[i].setRightNodeDesc(nodeRight);

        haarClassifierNodes.push_back(h_TmpClassifierNotRootNodes[i]);
    }

    return NCV_SUCCESS;
}

#endif /* HAVE_CUDA */