【特征匹配】ORB原理与源码解析

相关 :

   Fast原理与源码解析

   Brief描述子原理与源码解析

   Harris原理与源码解析

 

•  http://blog.csdn.net/luoshixian099/article/details/48523267

• ORB特征提取详解

• ORB特征点检测

 

  为了满足实时性的要求，前面文章中介绍过快速提取特征点算法Fast,以及特征描述子Brief。本篇文章介绍的ORB算法结合了Fast和Brief的速度优势，并做了改进，且ORB是免费。

   Ethan Rublee等人2011年在《ORB：An Efficient Alternative to SIFT or SURF》文章中提出了ORB算法。结合Fast与Brief算法，并给Fast特征点增加了方向性，使得特征点具有旋转不变性，并提出了构造金字塔方法，解决尺度不变性，但文章中没有具体详述。实验证明，ORB远优于之前的SIFT与SURF算法。

-------------------------------------------------------------------------------------------------------------------------------

论文核心内容概述：

1.构造金字塔，在每层金字塔上采用Fast算法提取特征点，采用Harris角点响应函数，按角点响应值排序，选取前N个特征点。

2. oFast:计算每个特征点的主方向，灰度质心法，计算特征点半径为r的圆形邻域范围内的灰度质心位置。从中心位置到质心位置的向量，定义为该特 征点的主方向。

  定义矩的计算公式，x,y∈[-r,r]：

                                 

             质心位置：

                               

               主方向：

                                  

3.rBrief:为了解决旋转不变性，把特征点的Patch旋转到主方向上(steered Brief)。通过实验得到，描述子在各个维度上的均值比较离散(偏离0.5)，同时维度间相关性很强，说明特征点描述子区分性不好，影响匹配的效果。论文中提出采取学习的方法，采用300K个训练样本点。每一个特征点，选取Patch大小为wp=31，Patch内每对点都采用wt=5大小的子窗口灰度均值做比较，子窗口的个数即为N=(wp-wt)*(wp-wt),从N个窗口中随机选两个做比较即构成描述子的一个bit,论文中采用M=205590种可能的情况：   

       ---------------------------------------------------------------------------------

        1.对所有样本点，做M种测试，构成M维的描述子，每个维度上非1即0；

        2.按均值对M个维度排序（以0.5为中心），组成向量T；

        3.贪婪搜索：把向量T中第一个元素移动到R中，然后继续取T的第二个元素，与R中的所有元素做相关性比较，如果相关性大于指定的阈值Threshold,           抛弃T的这个元素，否则加入到R中；

        4.重复第3个步骤，直到R中有256个元素，若检测完毕,少于256个元素，则降低阈值，重复上述步骤；

       ----------------------------------------------------------------------------------

    rBrief:通过上面的步骤取到的256对点，构成的描述子各维度间相关性很低，区分性好；

                                        

                                              训练前                                            训练后

---------------------------------------------------------------------------------------------------------------------------------

ORB算法步骤,参考opencv源码:

1.首先构造尺度金字塔；

   金字塔共n层，与SIFT不同，每层仅有一副图像；

   第s层的尺度为，Fator初始尺度(默认为1.2)，原图在第0层；

   第s层图像大小:

                              ；

2.在不同尺度上采用Fast检测特征点；在每一层上按公式计算需要提取的特征点数n,在本层上按Fast角点响应值排序，提取前2n个特征点，然后根据Harris   角点响应值排序， 取前n个特征点，作为本层的特征点；

3.计算每个特征点的主方向（质心法）；

4.旋转每个特征点的Patch到主方向，采用上述步骤3的选取的最优的256对特征点做τ测试，构成256维描述子，占32个字节；

                   ,,n=256

 

4.采用汉明距离做特征点匹配；

----------OpenCV源码解析-------------------------------------------------------

ORB类定义：位置..\features2d.hpp

nfeatures:需要的特征点总数；

scaleFactor:尺度因子；

nlevels:金字塔层数；

edgeThreshold:边界阈值；

firstLevel:起始层；

 WTA_K：描述子形成方法,WTA_K=2表示，采用两两比较；

 scoreType:角点响应函数，可以选择Harris或者Fast的方法；

 patchSize:特征点邻域大小；

    /*!
    ORB implementation.
    */
    class CV_EXPORTS_W ORB : public Feature2D
    {
    public:
    // the size of the signature in bytes
    enum { kBytes = 32, HARRIS_SCORE=0, FAST_SCORE=1 };
     
    CV_WRAP explicit ORB(int nfeatures = 500, float scaleFactor = 1.2f, int nlevels = 8, int edgeThreshold = 31,//构造函数
    int firstLevel = 0, int WTA_K=2, int scoreType=ORB::HARRIS_SCORE, int patchSize=31 );
     
    // returns the descriptor size in bytes
    int descriptorSize() const; //描述子占用的字节数，默认32字节
    // returns the descriptor type
    int descriptorType() const;//描述子类型，8位整形数
     
    // Compute the ORB features and descriptors on an image
    void operator()(InputArray image, InputArray mask, vector<KeyPoint>& keypoints) const;
     
    // Compute the ORB features and descriptors on an image
    void operator()( InputArray image, InputArray mask, vector<KeyPoint>& keypoints, //提取特征点与形成描述子
    OutputArray descriptors, bool useProvidedKeypoints=false ) const;
     
    AlgorithmInfo* info() const;
     
    protected:
     
    void computeImpl( const Mat& image, vector<KeyPoint>& keypoints, Mat& descriptors ) const;//计算描述子
    void detectImpl( const Mat& image, vector<KeyPoint>& keypoints, const Mat& mask=Mat() ) const;//检测特征点
     
    CV_PROP_RW int nfeatures;//特征点总数
    CV_PROP_RW double scaleFactor;//尺度因子
    CV_PROP_RW int nlevels;//金字塔内层数
    CV_PROP_RW int edgeThreshold;//边界阈值
    CV_PROP_RW int firstLevel;//开始层数
    CV_PROP_RW int WTA_K;//描述子形成方法，默认WTA_K=2，两两比较
    CV_PROP_RW int scoreType;//角点响应函数
    CV_PROP_RW int patchSize;//邻域Patch大小
    };

特征提取及形成描述子：通过这个函数对图像提取Fast特征点或者计算特征描述子

_image:输入图像；

_mask:掩码图像;

_keypoints:输入角点；

_descriptors:如果为空，只寻找特征点，不计算特征描述子；

_useProvidedKeypoints:如果为true,函数只计算特征描述子；

    /** Compute the ORB features and descriptors on an image
    * @param img the image to compute the features and descriptors on
    * @param mask the mask to apply
    * @param keypoints the resulting keypoints
    * @param descriptors the resulting descriptors
    * @param do_keypoints if true, the keypoints are computed, otherwise used as an input
    * @param do_descriptors if true, also computes the descriptors
    */
    void ORB::operator()( InputArray _image, InputArray _mask, vector<KeyPoint>& _keypoints,
    OutputArray _descriptors, bool useProvidedKeypoints) const
    {
    CV_Assert(patchSize >= 2);
     
    bool do_keypoints = !useProvidedKeypoints;
    bool do_descriptors = _descriptors.needed();
     
    if( (!do_keypoints && !do_descriptors) || _image.empty() )
    return;
     
    //ROI handling
    const int HARRIS_BLOCK_SIZE = 9;//Harris角点响应需要的边界大小
    int halfPatchSize = patchSize / 2;.//邻域半径
    int border = std::max(edgeThreshold, std::max(halfPatchSize, HARRIS_BLOCK_SIZE/2))+1;//采用最大的边界
     
    Mat image = _image.getMat(), mask = _mask.getMat();
    if( image.type() != CV_8UC1 )
    cvtColor(_image, image, CV_BGR2GRAY);//转灰度图
     
    int levelsNum = this->nlevels;//金字塔层数
     
    if( !do_keypoints ) //不做特征点检测
    {
    // if we have pre-computed keypoints, they may use more levels than it is set in parameters
    // !!!TODO!!! implement more correct method, independent from the used keypoint detector.
    // Namely, the detector should provide correct size of each keypoint. Based on the keypoint size
    // and the algorithm used (i.e. BRIEF, running on 31x31 patches) we should compute the approximate
    // scale-factor that we need to apply. Then we should cluster all the computed scale-factors and
    // for each cluster compute the corresponding image.
    //
    // In short, ultimately the descriptor should
    // ignore octave parameter and deal only with the keypoint size.
    levelsNum = 0;
    for( size_t i = 0; i < _keypoints.size(); i++ )
    levelsNum = std::max(levelsNum, std::max(_keypoints[i].octave, 0));//提取特征点的最大层数
    levelsNum++;
    }
     
    // Pre-compute the scale pyramids
    vector<Mat> imagePyramid(levelsNum), maskPyramid(levelsNum);//创建尺度金字塔图像
    for (int level = 0; level < levelsNum; ++level)
    {
    float scale = 1/getScale(level, firstLevel, scaleFactor); //每层对应的尺度
    /*
    static inline float getScale(int level, int firstLevel, double scaleFactor)
    {
    return (float)std::pow(scaleFactor, (double)(level - firstLevel));
    }
    */
    Size sz(cvRound(image.cols*scale), cvRound(image.rows*scale));//每层对应的图像大小
    Size wholeSize(sz.width + border*2, sz.height + border*2);
    Mat temp(wholeSize, image.type()), masktemp;
    imagePyramid[level] = temp(Rect(border, border, sz.width, sz.height));
    if( !mask.empty() )
    {
    masktemp = Mat(wholeSize, mask.type());
    maskPyramid[level] = masktemp(Rect(border, border, sz.width, sz.height));
    }
     
    // Compute the resized image
    if( level != firstLevel ) //得到金字塔每层的图像
    {
    if( level < firstLevel )
    {
    resize(image, imagePyramid[level], sz, 0, 0, INTER_LINEAR);
    if (!mask.empty())
    resize(mask, maskPyramid[level], sz, 0, 0, INTER_LINEAR);
    }
    else
    {
    resize(imagePyramid[level-1], imagePyramid[level], sz, 0, 0, INTER_LINEAR);
    if (!mask.empty())
    {
    resize(maskPyramid[level-1], maskPyramid[level], sz, 0, 0, INTER_LINEAR);
    threshold(maskPyramid[level], maskPyramid[level], 254, 0, THRESH_TOZERO);
    }
    }
     
    copyMakeBorder(imagePyramid[level], temp, border, border, border, border,//扩大图像的边界
    BORDER_REFLECT_101+BORDER_ISOLATED);
    if (!mask.empty())
    copyMakeBorder(maskPyramid[level], masktemp, border, border, border, border,
    BORDER_CONSTANT+BORDER_ISOLATED);
    }
    else
    {
    copyMakeBorder(image, temp, border, border, border, border,//扩大图像的四个边界
    BORDER_REFLECT_101);
    if( !mask.empty() )
    copyMakeBorder(mask, masktemp, border, border, border, border,
    BORDER_CONSTANT+BORDER_ISOLATED);
    }
    }
     
    // Pre-compute the keypoints (we keep the best over all scales, so this has to be done beforehand
    vector < vector<KeyPoint> > allKeypoints;
    if( do_keypoints )//提取角点
    {
    // Get keypoints, those will be far enough from the border that no check will be required for the descriptor
    computeKeyPoints(imagePyramid, maskPyramid, allKeypoints, //对每一层图像提取角点，见下面(1)的分析
    nfeatures, firstLevel, scaleFactor,
    edgeThreshold, patchSize, scoreType);
     
    // make sure we have the right number of keypoints keypoints
    /*vector<KeyPoint> temp;
     
    for (int level = 0; level < n_levels; ++level)
    {
    vector<KeyPoint>& keypoints = all_keypoints[level];
    temp.insert(temp.end(), keypoints.begin(), keypoints.end());
    keypoints.clear();
    }
     
    KeyPoint::retainBest(temp, n_features_);
     
    for (vector<KeyPoint>::iterator keypoint = temp.begin(),
    keypoint_end = temp.end(); keypoint != keypoint_end; ++keypoint)
    all_keypoints[keypoint->octave].push_back(*keypoint);*/
    }
    else //不提取角点
    {
    // Remove keypoints very close to the border
    KeyPointsFilter::runByImageBorder(_keypoints, image.size(), edgeThreshold);
     
    // Cluster the input keypoints depending on the level they were computed at
    allKeypoints.resize(levelsNum);
    for (vector<KeyPoint>::iterator keypoint = _keypoints.begin(),
    keypointEnd = _keypoints.end(); keypoint != keypointEnd; ++keypoint)
    allKeypoints[keypoint->octave].push_back(*keypoint); //把角点信息存入allKeypoints内
     
    // Make sure we rescale the coordinates
    for (int level = 0; level < levelsNum; ++level) //把角点位置信息缩放到指定层位置上
    {
    if (level == firstLevel)
    continue;
     
    vector<KeyPoint> & keypoints = allKeypoints[level];
    float scale = 1/getScale(level, firstLevel, scaleFactor);
    for (vector<KeyPoint>::iterator keypoint = keypoints.begin(),
    keypointEnd = keypoints.end(); keypoint != keypointEnd; ++keypoint)
    keypoint->pt *= scale; //缩放
    }
    }
     
    Mat descriptors;
    vector<Point> pattern;
     
    if( do_descriptors ) //计算特征描述子
    {
    int nkeypoints = 0;
    for (int level = 0; level < levelsNum; ++level)
    nkeypoints += (int)allKeypoints[level].size();//得到所有层的角点总数
    if( nkeypoints == 0 )
    _descriptors.release();
    else
    {
    _descriptors.create(nkeypoints, descriptorSize(), CV_8U);//创建一个矩阵存放描述子,每一行表示一个角点信息
    descriptors = _descriptors.getMat();
    }
     
    const int npoints = 512;//取512个点,共256对,产生256维描述子,32个字节
    Point patternbuf[npoints];
    const Point* pattern0 = (const Point*)bit_pattern_31_;//训练好的256对数据点位置
     
    if( patchSize != 31 )
    {
    pattern0 = patternbuf;
    makeRandomPattern(patchSize, patternbuf, npoints);
    }
     
    CV_Assert( WTA_K == 2 || WTA_K == 3 || WTA_K == 4 );
     
    if( WTA_K == 2 ) //WTA_K=2使用两个点之间作比较
    std::copy(pattern0, pattern0 + npoints, std::back_inserter(pattern));
    else
    {
    int ntuples = descriptorSize()*4;
    initializeOrbPattern(pattern0, pattern, ntuples, WTA_K, npoints);
    }
    }
     
    _keypoints.clear();
    int offset = 0;
    for (int level = 0; level < levelsNum; ++level)//依次计算每一层的角点描述子
    {
    // Get the features and compute their orientation
    vector<KeyPoint>& keypoints = allKeypoints[level];
    int nkeypoints = (int)keypoints.size();//本层内角点个数
     
    // Compute the descriptors
    if (do_descriptors)
    {
    Mat desc;
    if (!descriptors.empty())
    {
    desc = descriptors.rowRange(offset, offset + nkeypoints);
    }
     
    offset += nkeypoints; //偏移量
    // preprocess the resized image
    Mat& workingMat = imagePyramid[level];
    //boxFilter(working_mat, working_mat, working_mat.depth(), Size(5,5), Point(-1,-1), true, BORDER_REFLECT_101);
    GaussianBlur(workingMat, workingMat, Size(7, 7), 2, 2, BORDER_REFLECT_101);//高斯平滑图像
    computeDescriptors(workingMat, keypoints, desc, pattern, descriptorSize(), WTA_K);//计算本层内角点的描述子,（3）
    }
     
    // Copy to the output data
    if (level != firstLevel) //角点位置信息返回到原图上
    {
    float scale = getScale(level, firstLevel, scaleFactor);
    for (vector<KeyPoint>::iterator keypoint = keypoints.begin(),
    keypointEnd = keypoints.end(); keypoint != keypointEnd; ++keypoint)
    keypoint->pt *= scale;
    }
    // And add the keypoints to the output
    _keypoints.insert(_keypoints.end(), keypoints.begin(), keypoints.end());//存入描述子信息，返回
    }
    }

（1）提取角点：computeKeyPoints

imagePyramid:即构造好的金字塔

    /** Compute the ORB keypoints on an image
    * @param image_pyramid the image pyramid to compute the features and descriptors on
    * @param mask_pyramid the masks to apply at every level
    * @param keypoints the resulting keypoints, clustered per level
    */
    static void computeKeyPoints(const vector<Mat>& imagePyramid,
    const vector<Mat>& maskPyramid,
    vector<vector<KeyPoint> >& allKeypoints,
    int nfeatures, int firstLevel, double scaleFactor,
    int edgeThreshold, int patchSize, int scoreType )
    {
    int nlevels = (int)imagePyramid.size(); //金字塔层数
    vector<int> nfeaturesPerLevel(nlevels);
     
    // fill the extractors and descriptors for the corresponding scales
    float factor = (float)(1.0 / scaleFactor);
    float ndesiredFeaturesPerScale = nfeatures*(1 - factor)/(1 - (float)pow((double)factor, (double)nlevels));//
     
    int sumFeatures = 0;
    for( int level = 0; level < nlevels-1; level++ ) //对每层图像上分配相应角点数
    {
    nfeaturesPerLevel[level] = cvRound(ndesiredFeaturesPerScale);
    sumFeatures += nfeaturesPerLevel[level];
    ndesiredFeaturesPerScale *= factor;
    }
    nfeaturesPerLevel[nlevels-1] = std::max(nfeatures - sumFeatures, 0);//剩下角点数，由最上层图像提取
     
    // Make sure we forget about what is too close to the boundary
    //edge_threshold_ = std::max(edge_threshold_, patch_size_/2 + kKernelWidth / 2 + 2);
     
    // pre-compute the end of a row in a circular patch
    int halfPatchSize = patchSize / 2; //计算每个特征点圆邻域的位置信息
    vector<int> umax(halfPatchSize + 2);
    int v, v0, vmax = cvFloor(halfPatchSize * sqrt(2.f) / 2 + 1);
    int vmin = cvCeil(halfPatchSize * sqrt(2.f) / 2);
    for (v = 0; v <= vmax; ++v) //
    umax[v] = cvRound(sqrt((double)halfPatchSize * halfPatchSize - v * v));
    // Make sure we are symmetric
    for (v = halfPatchSize, v0 = 0; v >= vmin; --v)
    {
    while (umax[v0] == umax[v0 + 1])
    ++v0;
    umax[v] = v0;
    ++v0;
    }
     
    allKeypoints.resize(nlevels);
     
    for (int level = 0; level < nlevels; ++level)
    {
    int featuresNum = nfeaturesPerLevel[level];
    allKeypoints[level].reserve(featuresNum*2);
     
    vector<KeyPoint> & keypoints = allKeypoints[level];
     
    // Detect FAST features, 20 is a good threshold
    FastFeatureDetector fd(20, true);
    fd.detect(imagePyramid[level], keypoints, maskPyramid[level]);//Fast角点检测
     
    // Remove keypoints very close to the border
    KeyPointsFilter::runByImageBorder(keypoints, imagePyramid[level].size(), edgeThreshold);//去除邻近边界的点
     
    if( scoreType == ORB::HARRIS_SCORE )
    {
    // Keep more points than necessary as FAST does not give amazing corners
    KeyPointsFilter::retainBest(keypoints, 2 * featuresNum);//按Fast强度排序,保留前2*featuresNum个特征点
     
    // Compute the Harris cornerness (better scoring than FAST)
    HarrisResponses(imagePyramid[level], keypoints, 7, HARRIS_K); //计算每个角点的Harris强度响应
    }
     
    //cull to the final desired level, using the new Harris scores or the original FAST scores.
    KeyPointsFilter::retainBest(keypoints, featuresNum);//按Harris强度排序，保留前featuresNum个
     
    float sf = getScale(level, firstLevel, scaleFactor);
     
    // Set the level of the coordinates
    for (vector<KeyPoint>::iterator keypoint = keypoints.begin(),
    keypointEnd = keypoints.end(); keypoint != keypointEnd; ++keypoint)
    {
    keypoint->octave = level; //层信息
    keypoint->size = patchSize*sf; //
    }
     
    computeOrientation(imagePyramid[level], keypoints, halfPatchSize, umax); //计算角点的方向，（2）分析
    }
    }

（2）为每个角点计算主方向，质心法；

    static void computeOrientation(const Mat& image, vector<KeyPoint>& keypoints,
    int halfPatchSize, const vector<int>& umax)
    {
    // Process each keypoint
    for (vector<KeyPoint>::iterator keypoint = keypoints.begin(), //为每个角点计算主方向
    keypointEnd = keypoints.end(); keypoint != keypointEnd; ++keypoint)
    {
    keypoint->angle = IC_Angle(image, halfPatchSize, keypoint->pt, umax);//计算质心方向
    }
    }

    static float IC_Angle(const Mat& image, const int half_k, Point2f pt,
    const vector<int> & u_max)
    {
    int m_01 = 0, m_10 = 0;
     
    const uchar* center = &image.at<uchar> (cvRound(pt.y), cvRound(pt.x));
     
    // Treat the center line differently, v=0
    for (int u = -half_k; u <= half_k; ++u)
    m_10 += u * center[u];
     
    // Go line by line in the circular patch
    int step = (int)image.step1();
    for (int v = 1; v <= half_k; ++v) //每次处理对称的两行v
    {
    // Proceed over the two lines
    int v_sum = 0;
    int d = u_max[v];
    for (int u = -d; u <= d; ++u)
    {
    int val_plus = center[u + v*step], val_minus = center[u - v*step];
    v_sum += (val_plus - val_minus); //计算m_01时,位置上差一个符号
    m_10 += u * (val_plus + val_minus);
    }
    m_01 += v * v_sum;//计算上下两行的m_01
    }
     
    return fastAtan2((float)m_01, (float)m_10);//计算角度
    }

(3)计算特征点描述子

    static void computeDescriptors(const Mat& image, vector<KeyPoint>& keypoints, Mat& descriptors,
    const vector<Point>& pattern, int dsize, int WTA_K)
    {
    //convert to grayscale if more than one color
    CV_Assert(image.type() == CV_8UC1);
    //create the descriptor mat, keypoints.size() rows, BYTES cols
    descriptors = Mat::zeros((int)keypoints.size(), dsize, CV_8UC1);
     
    for (size_t i = 0; i < keypoints.size(); i++)
    computeOrbDescriptor(keypoints[i], image, &pattern[0], descriptors.ptr((int)i), dsize, WTA_K);
    }

        static void computeOrbDescriptor(const KeyPoint& kpt,
        const Mat& img, const Point* pattern,
        uchar* desc, int dsize, int WTA_K)
        {
        float angle = kpt.angle;
        //angle = cvFloor(angle/12)*12.f;
        angle *= (float)(CV_PI/180.f);
        float a = (float)cos(angle), b = (float)sin(angle);
        const uchar* center = &img.at<uchar>(cvRound(kpt.pt.y), cvRound(kpt.pt.x));
        int step = (int)img.step;
        #if 1
        #define GET_VALUE(idx) \ //取旋转后一个像素点的值
        center[cvRound(pattern[idx].x*b + pattern[idx].y*a)*step + \
        cvRound(pattern[idx].x*a - pattern[idx].y*b)]
        #else
        float x, y;
        int ix, iy;
        #define GET_VALUE(idx) \ //取旋转后一个像素点，插值法
        (x = pattern[idx].x*a - pattern[idx].y*b, \
        y = pattern[idx].x*b + pattern[idx].y*a, \
        ix = cvFloor(x), iy = cvFloor(y), \
        x -= ix, y -= iy, \
        cvRound(center[iy*step + ix]*(1-x)*(1-y) + center[(iy+1)*step + ix]*(1-x)*y + \
        center[iy*step + ix+1]*x*(1-y) + center[(iy+1)*step + ix+1]*x*y))
        #endif
        if( WTA_K == 2 )
        {
        for (int i = 0; i < dsize; ++i, pattern += 16)//每个特征描述子长度为32个字节
        {
        int t0, t1, val;
        t0 = GET_VALUE(0); t1 = GET_VALUE(1);
        val = t0 < t1;
        t0 = GET_VALUE(2); t1 = GET_VALUE(3);
        val |= (t0 < t1) << 1;
        t0 = GET_VALUE(4); t1 = GET_VALUE(5);
        val |= (t0 < t1) << 2;
        t0 = GET_VALUE(6); t1 = GET_VALUE(7);
        val |= (t0 < t1) << 3;
        t0 = GET_VALUE(8); t1 = GET_VALUE(9);
        val |= (t0 < t1) << 4;
        t0 = GET_VALUE(10); t1 = GET_VALUE(11);
        val |= (t0 < t1) << 5;
        t0 = GET_VALUE(12); t1 = GET_VALUE(13);
        val |= (t0 < t1) << 6;
        t0 = GET_VALUE(14); t1 = GET_VALUE(15);
        val |= (t0 < t1) << 7;
        desc[i] = (uchar)val;
        }
        }
        else if( WTA_K == 3 )
        {
        for (int i = 0; i < dsize; ++i, pattern += 12)
        {
        int t0, t1, t2, val;
        t0 = GET_VALUE(0); t1 = GET_VALUE(1); t2 = GET_VALUE(2);
        val = t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0);
        t0 = GET_VALUE(3); t1 = GET_VALUE(4); t2 = GET_VALUE(5);
        val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 2;
        t0 = GET_VALUE(6); t1 = GET_VALUE(7); t2 = GET_VALUE(8);
        val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 4;
        t0 = GET_VALUE(9); t1 = GET_VALUE(10); t2 = GET_VALUE(11);
        val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 6;
        desc[i] = (uchar)val;
        }
        }
        else if( WTA_K == 4 )
        {
        for (int i = 0; i < dsize; ++i, pattern += 16)
        {
        int t0, t1, t2, t3, u, v, k, val;
        t0 = GET_VALUE(0); t1 = GET_VALUE(1);
        t2 = GET_VALUE(2); t3 = GET_VALUE(3);
        u = 0, v = 2;
        if( t1 > t0 ) t0 = t1, u = 1;
        if( t3 > t2 ) t2 = t3, v = 3;
        k = t0 > t2 ? u : v;
        val = k;
        t0 = GET_VALUE(4); t1 = GET_VALUE(5);
        t2 = GET_VALUE(6); t3 = GET_VALUE(7);
        u = 0, v = 2;
        if( t1 > t0 ) t0 = t1, u = 1;
        if( t3 > t2 ) t2 = t3, v = 3;
        k = t0 > t2 ? u : v;
        val |= k << 2;
        t0 = GET_VALUE(8); t1 = GET_VALUE(9);
        t2 = GET_VALUE(10); t3 = GET_VALUE(11);
        u = 0, v = 2;
        if( t1 > t0 ) t0 = t1, u = 1;
        if( t3 > t2 ) t2 = t3, v = 3;
        k = t0 > t2 ? u : v;
        val |= k << 4;
        t0 = GET_VALUE(12); t1 = GET_VALUE(13);
        t2 = GET_VALUE(14); t3 = GET_VALUE(15);
        u = 0, v = 2;
        if( t1 > t0 ) t0 = t1, u = 1;
        if( t3 > t2 ) t2 = t3, v = 3;
        k = t0 > t2 ? u : v;
        val |= k << 6;
        desc[i] = (uchar)val;
        }
        }
        else
        CV_Error( CV_StsBadSize, "Wrong WTA_K. It can be only 2, 3 or 4." );
        #undef GET_VALUE
        }