Home | History | Annotate | Download | only in src
      1 /*M///////////////////////////////////////////////////////////////////////////////////////
      2 //
      3 //  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
      4 //
      5 //  By downloading, copying, installing or using the software you agree to this license.
      6 //  If you do not agree to this license, do not download, install,
      7 //  copy or use the software.
      8 //
      9 //
     10 //                          License Agreement
     11 //                For Open Source Computer Vision Library
     12 //
     13 // Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
     14 // Copyright (C) 2009, Willow Garage Inc., all rights reserved.
     15 // Copyright (C) 2013, OpenCV Foundation, all rights reserved.
     16 // Third party copyrights are property of their respective owners.
     17 //
     18 // Redistribution and use in source and binary forms, with or without modification,
     19 // are permitted provided that the following conditions are met:
     20 //
     21 //   * Redistribution's of source code must retain the above copyright notice,
     22 //     this list of conditions and the following disclaimer.
     23 //
     24 //   * Redistribution's in binary form must reproduce the above copyright notice,
     25 //     this list of conditions and the following disclaimer in the documentation
     26 //     and/or other materials provided with the distribution.
     27 //
     28 //   * The name of the copyright holders may not be used to endorse or promote products
     29 //     derived from this software without specific prior written permission.
     30 //
     31 // This software is provided by the copyright holders and contributors "as is" and
     32 // any express or implied warranties, including, but not limited to, the implied
     33 // warranties of merchantability and fitness for a particular purpose are disclaimed.
     34 // In no event shall the Intel Corporation or contributors be liable for any direct,
     35 // indirect, incidental, special, exemplary, or consequential damages
     36 // (including, but not limited to, procurement of substitute goods or services;
     37 // loss of use, data, or profits; or business interruption) however caused
     38 // and on any theory of liability, whether in contract, strict liability,
     39 // or tort (including negligence or otherwise) arising in any way out of
     40 // the use of this software, even if advised of the possibility of such damage.
     41 //
     42 //M*/
     43 
     44 #include "precomp.hpp"
     45 
     46 ////////////////////////////////////////// kmeans ////////////////////////////////////////////
     47 
     48 namespace cv
     49 {
     50 
     51 static void generateRandomCenter(const std::vector<Vec2f>& box, float* center, RNG& rng)
     52 {
     53     size_t j, dims = box.size();
     54     float margin = 1.f/dims;
     55     for( j = 0; j < dims; j++ )
     56         center[j] = ((float)rng*(1.f+margin*2.f)-margin)*(box[j][1] - box[j][0]) + box[j][0];
     57 }
     58 
     59 class KMeansPPDistanceComputer : public ParallelLoopBody
     60 {
     61 public:
     62     KMeansPPDistanceComputer( float *_tdist2,
     63                               const float *_data,
     64                               const float *_dist,
     65                               int _dims,
     66                               size_t _step,
     67                               size_t _stepci )
     68         : tdist2(_tdist2),
     69           data(_data),
     70           dist(_dist),
     71           dims(_dims),
     72           step(_step),
     73           stepci(_stepci) { }
     74 
     75     void operator()( const cv::Range& range ) const
     76     {
     77         const int begin = range.start;
     78         const int end = range.end;
     79 
     80         for ( int i = begin; i<end; i++ )
     81         {
     82             tdist2[i] = std::min(normL2Sqr(data + step*i, data + stepci, dims), dist[i]);
     83         }
     84     }
     85 
     86 private:
     87     KMeansPPDistanceComputer& operator=(const KMeansPPDistanceComputer&); // to quiet MSVC
     88 
     89     float *tdist2;
     90     const float *data;
     91     const float *dist;
     92     const int dims;
     93     const size_t step;
     94     const size_t stepci;
     95 };
     96 
     97 /*
     98 k-means center initialization using the following algorithm:
     99 Arthur & Vassilvitskii (2007) k-means++: The Advantages of Careful Seeding
    100 */
    101 static void generateCentersPP(const Mat& _data, Mat& _out_centers,
    102                               int K, RNG& rng, int trials)
    103 {
    104     int i, j, k, dims = _data.cols, N = _data.rows;
    105     const float* data = _data.ptr<float>(0);
    106     size_t step = _data.step/sizeof(data[0]);
    107     std::vector<int> _centers(K);
    108     int* centers = &_centers[0];
    109     std::vector<float> _dist(N*3);
    110     float* dist = &_dist[0], *tdist = dist + N, *tdist2 = tdist + N;
    111     double sum0 = 0;
    112 
    113     centers[0] = (unsigned)rng % N;
    114 
    115     for( i = 0; i < N; i++ )
    116     {
    117         dist[i] = normL2Sqr(data + step*i, data + step*centers[0], dims);
    118         sum0 += dist[i];
    119     }
    120 
    121     for( k = 1; k < K; k++ )
    122     {
    123         double bestSum = DBL_MAX;
    124         int bestCenter = -1;
    125 
    126         for( j = 0; j < trials; j++ )
    127         {
    128             double p = (double)rng*sum0, s = 0;
    129             for( i = 0; i < N-1; i++ )
    130                 if( (p -= dist[i]) <= 0 )
    131                     break;
    132             int ci = i;
    133 
    134             parallel_for_(Range(0, N),
    135                          KMeansPPDistanceComputer(tdist2, data, dist, dims, step, step*ci));
    136             for( i = 0; i < N; i++ )
    137             {
    138                 s += tdist2[i];
    139             }
    140 
    141             if( s < bestSum )
    142             {
    143                 bestSum = s;
    144                 bestCenter = ci;
    145                 std::swap(tdist, tdist2);
    146             }
    147         }
    148         centers[k] = bestCenter;
    149         sum0 = bestSum;
    150         std::swap(dist, tdist);
    151     }
    152 
    153     for( k = 0; k < K; k++ )
    154     {
    155         const float* src = data + step*centers[k];
    156         float* dst = _out_centers.ptr<float>(k);
    157         for( j = 0; j < dims; j++ )
    158             dst[j] = src[j];
    159     }
    160 }
    161 
    162 class KMeansDistanceComputer : public ParallelLoopBody
    163 {
    164 public:
    165     KMeansDistanceComputer( double *_distances,
    166                             int *_labels,
    167                             const Mat& _data,
    168                             const Mat& _centers )
    169         : distances(_distances),
    170           labels(_labels),
    171           data(_data),
    172           centers(_centers)
    173     {
    174     }
    175 
    176     void operator()( const Range& range ) const
    177     {
    178         const int begin = range.start;
    179         const int end = range.end;
    180         const int K = centers.rows;
    181         const int dims = centers.cols;
    182 
    183         for( int i = begin; i<end; ++i)
    184         {
    185             const float *sample = data.ptr<float>(i);
    186             int k_best = 0;
    187             double min_dist = DBL_MAX;
    188 
    189             for( int k = 0; k < K; k++ )
    190             {
    191                 const float* center = centers.ptr<float>(k);
    192                 const double dist = normL2Sqr(sample, center, dims);
    193 
    194                 if( min_dist > dist )
    195                 {
    196                     min_dist = dist;
    197                     k_best = k;
    198                 }
    199             }
    200 
    201             distances[i] = min_dist;
    202             labels[i] = k_best;
    203         }
    204     }
    205 
    206 private:
    207     KMeansDistanceComputer& operator=(const KMeansDistanceComputer&); // to quiet MSVC
    208 
    209     double *distances;
    210     int *labels;
    211     const Mat& data;
    212     const Mat& centers;
    213 };
    214 
    215 }
    216 
    217 double cv::kmeans( InputArray _data, int K,
    218                    InputOutputArray _bestLabels,
    219                    TermCriteria criteria, int attempts,
    220                    int flags, OutputArray _centers )
    221 {
    222     const int SPP_TRIALS = 3;
    223     Mat data0 = _data.getMat();
    224     bool isrow = data0.rows == 1 && data0.channels() > 1;
    225     int N = !isrow ? data0.rows : data0.cols;
    226     int dims = (!isrow ? data0.cols : 1)*data0.channels();
    227     int type = data0.depth();
    228 
    229     attempts = std::max(attempts, 1);
    230     CV_Assert( data0.dims <= 2 && type == CV_32F && K > 0 );
    231     CV_Assert( N >= K );
    232 
    233     Mat data(N, dims, CV_32F, data0.ptr(), isrow ? dims * sizeof(float) : static_cast<size_t>(data0.step));
    234 
    235     _bestLabels.create(N, 1, CV_32S, -1, true);
    236 
    237     Mat _labels, best_labels = _bestLabels.getMat();
    238     if( flags & CV_KMEANS_USE_INITIAL_LABELS )
    239     {
    240         CV_Assert( (best_labels.cols == 1 || best_labels.rows == 1) &&
    241                   best_labels.cols*best_labels.rows == N &&
    242                   best_labels.type() == CV_32S &&
    243                   best_labels.isContinuous());
    244         best_labels.copyTo(_labels);
    245     }
    246     else
    247     {
    248         if( !((best_labels.cols == 1 || best_labels.rows == 1) &&
    249              best_labels.cols*best_labels.rows == N &&
    250             best_labels.type() == CV_32S &&
    251             best_labels.isContinuous()))
    252             best_labels.create(N, 1, CV_32S);
    253         _labels.create(best_labels.size(), best_labels.type());
    254     }
    255     int* labels = _labels.ptr<int>();
    256 
    257     Mat centers(K, dims, type), old_centers(K, dims, type), temp(1, dims, type);
    258     std::vector<int> counters(K);
    259     std::vector<Vec2f> _box(dims);
    260     Vec2f* box = &_box[0];
    261     double best_compactness = DBL_MAX, compactness = 0;
    262     RNG& rng = theRNG();
    263     int a, iter, i, j, k;
    264 
    265     if( criteria.type & TermCriteria::EPS )
    266         criteria.epsilon = std::max(criteria.epsilon, 0.);
    267     else
    268         criteria.epsilon = FLT_EPSILON;
    269     criteria.epsilon *= criteria.epsilon;
    270 
    271     if( criteria.type & TermCriteria::COUNT )
    272         criteria.maxCount = std::min(std::max(criteria.maxCount, 2), 100);
    273     else
    274         criteria.maxCount = 100;
    275 
    276     if( K == 1 )
    277     {
    278         attempts = 1;
    279         criteria.maxCount = 2;
    280     }
    281 
    282     const float* sample = data.ptr<float>(0);
    283     for( j = 0; j < dims; j++ )
    284         box[j] = Vec2f(sample[j], sample[j]);
    285 
    286     for( i = 1; i < N; i++ )
    287     {
    288         sample = data.ptr<float>(i);
    289         for( j = 0; j < dims; j++ )
    290         {
    291             float v = sample[j];
    292             box[j][0] = std::min(box[j][0], v);
    293             box[j][1] = std::max(box[j][1], v);
    294         }
    295     }
    296 
    297     for( a = 0; a < attempts; a++ )
    298     {
    299         double max_center_shift = DBL_MAX;
    300         for( iter = 0;; )
    301         {
    302             swap(centers, old_centers);
    303 
    304             if( iter == 0 && (a > 0 || !(flags & KMEANS_USE_INITIAL_LABELS)) )
    305             {
    306                 if( flags & KMEANS_PP_CENTERS )
    307                     generateCentersPP(data, centers, K, rng, SPP_TRIALS);
    308                 else
    309                 {
    310                     for( k = 0; k < K; k++ )
    311                         generateRandomCenter(_box, centers.ptr<float>(k), rng);
    312                 }
    313             }
    314             else
    315             {
    316                 if( iter == 0 && a == 0 && (flags & KMEANS_USE_INITIAL_LABELS) )
    317                 {
    318                     for( i = 0; i < N; i++ )
    319                         CV_Assert( (unsigned)labels[i] < (unsigned)K );
    320                 }
    321 
    322                 // compute centers
    323                 centers = Scalar(0);
    324                 for( k = 0; k < K; k++ )
    325                     counters[k] = 0;
    326 
    327                 for( i = 0; i < N; i++ )
    328                 {
    329                     sample = data.ptr<float>(i);
    330                     k = labels[i];
    331                     float* center = centers.ptr<float>(k);
    332                     j=0;
    333                     #if CV_ENABLE_UNROLLED
    334                     for(; j <= dims - 4; j += 4 )
    335                     {
    336                         float t0 = center[j] + sample[j];
    337                         float t1 = center[j+1] + sample[j+1];
    338 
    339                         center[j] = t0;
    340                         center[j+1] = t1;
    341 
    342                         t0 = center[j+2] + sample[j+2];
    343                         t1 = center[j+3] + sample[j+3];
    344 
    345                         center[j+2] = t0;
    346                         center[j+3] = t1;
    347                     }
    348                     #endif
    349                     for( ; j < dims; j++ )
    350                         center[j] += sample[j];
    351                     counters[k]++;
    352                 }
    353 
    354                 if( iter > 0 )
    355                     max_center_shift = 0;
    356 
    357                 for( k = 0; k < K; k++ )
    358                 {
    359                     if( counters[k] != 0 )
    360                         continue;
    361 
    362                     // if some cluster appeared to be empty then:
    363                     //   1. find the biggest cluster
    364                     //   2. find the farthest from the center point in the biggest cluster
    365                     //   3. exclude the farthest point from the biggest cluster and form a new 1-point cluster.
    366                     int max_k = 0;
    367                     for( int k1 = 1; k1 < K; k1++ )
    368                     {
    369                         if( counters[max_k] < counters[k1] )
    370                             max_k = k1;
    371                     }
    372 
    373                     double max_dist = 0;
    374                     int farthest_i = -1;
    375                     float* new_center = centers.ptr<float>(k);
    376                     float* old_center = centers.ptr<float>(max_k);
    377                     float* _old_center = temp.ptr<float>(); // normalized
    378                     float scale = 1.f/counters[max_k];
    379                     for( j = 0; j < dims; j++ )
    380                         _old_center[j] = old_center[j]*scale;
    381 
    382                     for( i = 0; i < N; i++ )
    383                     {
    384                         if( labels[i] != max_k )
    385                             continue;
    386                         sample = data.ptr<float>(i);
    387                         double dist = normL2Sqr(sample, _old_center, dims);
    388 
    389                         if( max_dist <= dist )
    390                         {
    391                             max_dist = dist;
    392                             farthest_i = i;
    393                         }
    394                     }
    395 
    396                     counters[max_k]--;
    397                     counters[k]++;
    398                     labels[farthest_i] = k;
    399                     sample = data.ptr<float>(farthest_i);
    400 
    401                     for( j = 0; j < dims; j++ )
    402                     {
    403                         old_center[j] -= sample[j];
    404                         new_center[j] += sample[j];
    405                     }
    406                 }
    407 
    408                 for( k = 0; k < K; k++ )
    409                 {
    410                     float* center = centers.ptr<float>(k);
    411                     CV_Assert( counters[k] != 0 );
    412 
    413                     float scale = 1.f/counters[k];
    414                     for( j = 0; j < dims; j++ )
    415                         center[j] *= scale;
    416 
    417                     if( iter > 0 )
    418                     {
    419                         double dist = 0;
    420                         const float* old_center = old_centers.ptr<float>(k);
    421                         for( j = 0; j < dims; j++ )
    422                         {
    423                             double t = center[j] - old_center[j];
    424                             dist += t*t;
    425                         }
    426                         max_center_shift = std::max(max_center_shift, dist);
    427                     }
    428                 }
    429             }
    430 
    431             if( ++iter == MAX(criteria.maxCount, 2) || max_center_shift <= criteria.epsilon )
    432                 break;
    433 
    434             // assign labels
    435             Mat dists(1, N, CV_64F);
    436             double* dist = dists.ptr<double>(0);
    437             parallel_for_(Range(0, N),
    438                          KMeansDistanceComputer(dist, labels, data, centers));
    439             compactness = 0;
    440             for( i = 0; i < N; i++ )
    441             {
    442                 compactness += dist[i];
    443             }
    444         }
    445 
    446         if( compactness < best_compactness )
    447         {
    448             best_compactness = compactness;
    449             if( _centers.needed() )
    450                 centers.copyTo(_centers);
    451             _labels.copyTo(best_labels);
    452         }
    453     }
    454 
    455     return best_compactness;
    456 }
    457