fixed 2228

8d73bbb8 · marina.kolpakova · 989631c5 · 8d73bbb8
Commit 8d73bbb8 authored Aug 02, 2012 by marina.kolpakova
Hide whitespace changes
Inline Side-by-side

Showing with 469 additions and 467 deletions

cascadeclassifier.cpp modules/gpu/src/cascadeclassifier.cpp +469 -467

No files found.
--- a/modules/gpu/src/cascadeclassifier.cpp
+++ b/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)
@@ -94,219 +94,221 @@ public:
                      /*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;
-    }
+        lastAllocatedFrameSize = frameSize;
+        return NCV_SUCCESS;
-    cudaDeviceProp devProp;
+    }
-    NCVStatus ncvStat;
+    cudaDeviceProp devProp;
-    Ptr<NCVMemNativeAllocator> gpuCascadeAllocator;
+    NCVStatus ncvStat;
-    Ptr<NCVMemNativeAllocator> cpuCascadeAllocator;
+    Ptr<NCVMemNativeAllocator> gpuCascadeAllocator;
-    Ptr<NCVVectorAlloc<HaarStage64> >           h_haarStages;
+    Ptr<NCVMemNativeAllocator> cpuCascadeAllocator;
-    Ptr<NCVVectorAlloc<HaarClassifierNode128> > h_haarNodes;
-    Ptr<NCVVectorAlloc<HaarFeature64> >         h_haarFeatures;
+    Ptr<NCVVectorAlloc<HaarStage64> >           h_haarStages;
+    Ptr<NCVVectorAlloc<HaarClassifierNode128> > h_haarNodes;
-    HaarClassifierCascadeDescriptor haar;
+    Ptr<NCVVectorAlloc<HaarFeature64> >         h_haarFeatures;
-    Ptr<NCVVectorAlloc<HaarStage64> >           d_haarStages;
+    HaarClassifierCascadeDescriptor haar;
-    Ptr<NCVVectorAlloc<HaarClassifierNode128> > d_haarNodes;
-    Ptr<NCVVectorAlloc<HaarFeature64> >         d_haarFeatures;
+    Ptr<NCVVectorAlloc<HaarStage64> >           d_haarStages;
+    Ptr<NCVVectorAlloc<HaarClassifierNode128> > d_haarNodes;
-    Size lastAllocatedFrameSize;
+    Ptr<NCVVectorAlloc<HaarFeature64> >         d_haarFeatures;
-    Ptr<NCVMemStackAllocator> gpuAllocator;
+    Size lastAllocatedFrameSize;
-    Ptr<NCVMemStackAllocator> cpuAllocator;
+    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)
@@ -669,18 +671,18 @@ 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());
@@ -695,261 +697,261 @@ bool cv::gpu::CascadeClassifier_GPU::load(const string& filename)
    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 */