void CvEM::kmeans( const CvVectors& train_data, int nclusters, CvMat* labels,
CvTermCriteria termcrit, const CvMat* centers0 )
{
CvMat* centers = 0;
CvMat* old_centers = 0;
CvMat* counters = 0;
CV_FUNCNAME( "CvEM::kmeans" );
__BEGIN__;
CvRNG rng = cvRNG(-1);
int i, j, k, nsamples, dims;
int iter = 0;
double max_dist = DBL_MAX;
termcrit = cvCheckTermCriteria( termcrit, 1e-6, 100 );
termcrit.epsilon *= termcrit.epsilon;
nsamples = train_data.count;
dims = train_data.dims;
nclusters = MIN( nclusters, nsamples );
CV_CALL( centers = cvCreateMat( nclusters, dims, CV_64FC1 ));
CV_CALL( old_centers = cvCreateMat( nclusters, dims, CV_64FC1 ));
CV_CALL( counters = cvCreateMat( 1, nclusters, CV_32SC1 ));
cvZero( old_centers );
if( centers0 )
{
CV_CALL( cvConvert( centers0, centers ));
}
else
{
for( i = 0; i < nsamples; i++ )
labels->data.i[i] = i*nclusters/nsamples;
cvRandShuffle( labels, &rng );
}
for( ;; )
{
CvMat* temp;
if( iter > 0 || centers0 )
{
for( i = 0; i < nsamples; i++ )
{
const float* s = train_data.data.fl[i];
int k_best = 0;
double min_dist = DBL_MAX;
for( k = 0; k < nclusters; k++ )
{
const double* c = (double*)(centers->data.ptr + k*centers->step);
double dist = 0;
for( j = 0; j <= dims - 4; j += 4 )
{
double t0 = c[j] - s[j];
double t1 = c[j+1] - s[j+1];
dist += t0*t0 + t1*t1;
t0 = c[j+2] - s[j+2];
t1 = c[j+3] - s[j+3];
dist += t0*t0 + t1*t1;
}
for( ; j < dims; j++ )
{
double t = c[j] - s[j];
dist += t*t;
}
if( min_dist > dist )
{
min_dist = dist;
k_best = k;
}
}
labels->data.i[i] = k_best;
}
}
if( ++iter > termcrit.max_iter )
break;
CV_SWAP( centers, old_centers, temp );
cvZero( centers );
cvZero( counters );
// update centers
for( i = 0; i < nsamples; i++ )
{
const float* s = train_data.data.fl[i];
k = labels->data.i[i];
double* c = (double*)(centers->data.ptr + k*centers->step);
for( j = 0; j <= dims - 4; j += 4 )
{
double t0 = c[j] + s[j];
double t1 = c[j+1] + s[j+1];
c[j] = t0;
c[j+1] = t1;
t0 = c[j+2] + s[j+2];
t1 = c[j+3] + s[j+3];
c[j+2] = t0;
c[j+3] = t1;
}
for( ; j < dims; j++ )
c[j] += s[j];
counters->data.i[k]++;
}
if( iter > 1 )
max_dist = 0;
for( k = 0; k < nclusters; k++ )
{
double* c = (double*)(centers->data.ptr + k*centers->step);
if( counters->data.i[k] != 0 )
{
double scale = 1./counters->data.i[k];
for( j = 0; j < dims; j++ )
c[j] *= scale;
}
else
{
const float* s;
for( j = 0; j < 10; j++ )
{
i = cvRandInt( &rng ) % nsamples;
if( counters->data.i[labels->data.i[i]] > 1 )
break;
}
s = train_data.data.fl[i];
for( j = 0; j < dims; j++ )
c[j] = s[j];
}
if( iter > 1 )
{
double dist = 0;
const double* c_o = (double*)(old_centers->data.ptr + k*old_centers->step);
for( j = 0; j < dims; j++ )
{
double t = c[j] - c_o[j];
dist += t*t;
}
if( max_dist < dist )
max_dist = dist;
}
}
if( max_dist < termcrit.epsilon )
break;
}
cvZero( counters );
for( i = 0; i < nsamples; i++ )
counters->data.i[labels->data.i[i]]++;
// ensure that we do not have empty clusters
for( k = 0; k < nclusters; k++ )
if( counters->data.i[k] == 0 )
for(;;)
{
i = cvRandInt(&rng) % nsamples;
j = labels->data.i[i];
if( counters->data.i[j] > 1 )
{
labels->data.i[i] = k;
counters->data.i[j]--;
counters->data.i[k]++;
break;
}
}
__END__;
cvReleaseMat( ¢ers );
cvReleaseMat( &old_centers );
cvReleaseMat( &counters );
}
void CvEM::set_params( const CvEMParams& _params, const CvVectors& train_data )
{
CV_FUNCNAME( "CvEM::set_params" );
__BEGIN__;
int k;
params = _params;
params.term_crit = cvCheckTermCriteria( params.term_crit, 1e-6, 10000 );
if( params.cov_mat_type != COV_MAT_SPHERICAL &&
params.cov_mat_type != COV_MAT_DIAGONAL &&
params.cov_mat_type != COV_MAT_GENERIC )
CV_ERROR( CV_StsBadArg, "Unknown covariation matrix type" );
switch( params.start_step )
{
case START_M_STEP:
if( !params.probs )
CV_ERROR( CV_StsNullPtr, "Probabilities must be specified when EM algorithm starts with M-step" );
break;
case START_E_STEP:
if( !params.means )
CV_ERROR( CV_StsNullPtr, "Mean's must be specified when EM algorithm starts with E-step" );
break;
case START_AUTO_STEP:
break;
default:
CV_ERROR( CV_StsBadArg, "Unknown start_step" );
}
if( params.nclusters < 1 )
CV_ERROR( CV_StsOutOfRange, "The number of clusters (mixtures) should be > 0" );
if( params.probs )
{
const CvMat* p = params.weights;
if( !CV_IS_MAT(p) ||
(CV_MAT_TYPE(p->type) != CV_32FC1 &&
CV_MAT_TYPE(p->type) != CV_64FC1) ||
p->rows != train_data.count ||
p->cols != params.nclusters )
CV_ERROR( CV_StsBadArg, "The array of probabilities must be a valid "
"floating-point matrix (CvMat) of 'nsamples' x 'nclusters' size" );
}
if( params.means )
{
const CvMat* m = params.means;
if( !CV_IS_MAT(m) ||
(CV_MAT_TYPE(m->type) != CV_32FC1 &&
CV_MAT_TYPE(m->type) != CV_64FC1) ||
m->rows != params.nclusters ||
m->cols != train_data.dims )
CV_ERROR( CV_StsBadArg, "The array of mean's must be a valid "
"floating-point matrix (CvMat) of 'nsamples' x 'dims' size" );
}
if( params.weights )
{
const CvMat* w = params.weights;
if( !CV_IS_MAT(w) ||
(CV_MAT_TYPE(w->type) != CV_32FC1 &&
CV_MAT_TYPE(w->type) != CV_64FC1) ||
(w->rows != 1 && w->cols != 1) ||
w->rows + w->cols - 1 != params.nclusters )
CV_ERROR( CV_StsBadArg, "The array of weights must be a valid "
"1d floating-point vector (CvMat) of 'nclusters' elements" );
}
if( params.covs )
for( k = 0; k < params.nclusters; k++ )
{
const CvMat* cov = params.covs[k];
if( !CV_IS_MAT(cov) ||
(CV_MAT_TYPE(cov->type) != CV_32FC1 &&
CV_MAT_TYPE(cov->type) != CV_64FC1) ||
cov->rows != cov->cols || cov->cols != train_data.dims )
CV_ERROR( CV_StsBadArg,
"Each of covariation matrices must be a valid square "
"floating-point matrix (CvMat) of 'dims' x 'dims'" );
}
__END__;
}
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: cvMeanShift
// Purpose: MeanShift algorithm
// Context:
// Parameters:
// imgProb - 2D object probability distribution
// windowIn - CvRect of CAMSHIFT Window intial size
// numIters - If CAMSHIFT iterates this many times, stop
// windowOut - Location, height and width of converged CAMSHIFT window
// len - If != NULL, return equivalent len
// width - If != NULL, return equivalent width
// itersUsed - Returns number of iterations CAMSHIFT took to converge
// Returns:
// The function itself returns the area found
// Notes:
//F*/
CV_IMPL int
cvMeanShift( const void* imgProb, CvRect windowIn,
CvTermCriteria criteria, CvConnectedComp* comp )
{
CvMoments moments;
int i = 0, eps;
CvMat stub, *mat = (CvMat*)imgProb;
CvMat cur_win;
CvRect cur_rect = windowIn;
CV_FUNCNAME( "cvMeanShift" );
if( comp )
comp->rect = windowIn;
moments.m00 = moments.m10 = moments.m01 = 0;
__BEGIN__;
CV_CALL( mat = cvGetMat( mat, &stub ));
if( CV_MAT_CN( mat->type ) > 1 )
CV_ERROR( CV_BadNumChannels, cvUnsupportedFormat );
if( windowIn.height <= 0 || windowIn.width <= 0 )
CV_ERROR( CV_StsBadArg, "Input window has non-positive sizes" );
if( windowIn.x < 0 || windowIn.x + windowIn.width > mat->cols ||
windowIn.y < 0 || windowIn.y + windowIn.height > mat->rows )
CV_ERROR( CV_StsBadArg, "Initial window is not inside the image ROI" );
CV_CALL( criteria = cvCheckTermCriteria( criteria, 1., 100 ));
eps = cvRound( criteria.epsilon * criteria.epsilon );
for( i = 0; i < criteria.max_iter; i++ )
{
int dx, dy, nx, ny;
double inv_m00;
CV_CALL( cvGetSubRect( mat, &cur_win, cur_rect ));
CV_CALL( cvMoments( &cur_win, &moments ));
/* Calculating center of mass */
if( fabs(moments.m00) < DBL_EPSILON )
break;
inv_m00 = moments.inv_sqrt_m00*moments.inv_sqrt_m00;
dx = cvRound( moments.m10 * inv_m00 - windowIn.width*0.5 );
dy = cvRound( moments.m01 * inv_m00 - windowIn.height*0.5 );
nx = cur_rect.x + dx;
ny = cur_rect.y + dy;
if( nx < 0 )
nx = 0;
else if( nx + cur_rect.width > mat->cols )
nx = mat->cols - cur_rect.width;
if( ny < 0 )
ny = 0;
else if( ny + cur_rect.height > mat->rows )
ny = mat->rows - cur_rect.height;
dx = nx - cur_rect.x;
dy = ny - cur_rect.y;
cur_rect.x = nx;
cur_rect.y = ny;
/* Check for coverage centers mass & window */
if( dx*dx + dy*dy < eps )
break;
}
__END__;
if( comp )
{
comp->rect = cur_rect;
comp->area = (float)moments.m00;
}
return i;
}
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: icvContourFromContourTree
// Purpose:
// reconstracts contour from binary tree representation
// Context:
// Parameters:
// tree - pointer to the input binary tree representation
// storage - pointer to the current storage block
// contour - pointer to output contour object.
// criteria - criteria for the definition threshold value
// for the contour reconstracting (level or precision)
//F*/
CV_IMPL CvSeq*
cvContourFromContourTree( const CvContourTree* tree,
CvMemStorage* storage,
CvTermCriteria criteria )
{
CvSeq* contour = 0;
_CvTrianAttr **ptr_buf = 0; /* pointer to the pointer's buffer */
int *level_buf = 0;
int i_buf;
int lpt;
double area_all;
double threshold;
int cur_level;
int level;
int seq_flags;
char log_iter, log_eps;
int out_hearder_size;
_CvTrianAttr *tree_one = 0, tree_root; /* current vertex */
CvSeqReader reader;
CvSeqWriter writer;
CV_FUNCNAME("cvContourFromContourTree");
__BEGIN__;
if( !tree )
CV_ERROR( CV_StsNullPtr, "" );
if( !CV_IS_SEQ_POLYGON_TREE( tree ))
CV_ERROR_FROM_STATUS( CV_BADFLAG_ERR );
criteria = cvCheckTermCriteria( criteria, 0., 100 );
lpt = tree->total;
ptr_buf = NULL;
level_buf = NULL;
i_buf = 0;
cur_level = 0;
log_iter = (char) (criteria.type == CV_TERMCRIT_ITER ||
(criteria.type == CV_TERMCRIT_ITER + CV_TERMCRIT_EPS));
log_eps = (char) (criteria.type == CV_TERMCRIT_EPS ||
(criteria.type == CV_TERMCRIT_ITER + CV_TERMCRIT_EPS));
cvStartReadSeq( (CvSeq *) tree, &reader, 0 );
out_hearder_size = sizeof( CvContour );
seq_flags = CV_SEQ_POLYGON;
cvStartWriteSeq( seq_flags, out_hearder_size, sizeof( CvPoint ), storage, &writer );
ptr_buf = (_CvTrianAttr **) cvAlloc( lpt * sizeof( _CvTrianAttr * ));
if( ptr_buf == NULL )
CV_ERROR_FROM_STATUS( CV_OUTOFMEM_ERR );
if( log_iter )
{
level_buf = (int *) cvAlloc( lpt * (sizeof( int )));
if( level_buf == NULL )
CV_ERROR_FROM_STATUS( CV_OUTOFMEM_ERR );
}
memset( ptr_buf, 0, lpt * sizeof( _CvTrianAttr * ));
/* write the first tree root's point as a start point of the result contour */
CV_WRITE_SEQ_ELEM( tree->p1, writer );
/* write the second tree root"s point into buffer */
/* read the root of the tree */
CV_READ_SEQ_ELEM( tree_root, reader );
tree_one = &tree_root;
area_all = tree_one->area;
if( log_eps )
threshold = criteria.epsilon * area_all;
else
threshold = 10 * area_all;
if( log_iter )
level = criteria.max_iter;
else
level = -1;
/* contour from binary tree constraction */
while( i_buf >= 0 )
{
if( tree_one != NULL && (cur_level <= level || tree_one->area >= threshold) )
/* go to left sub tree for the vertex and save pointer to the right vertex */
/* into the buffer */
{
ptr_buf[i_buf] = tree_one;
if( log_iter )
{
level_buf[i_buf] = cur_level;
cur_level++;
}
i_buf++;
tree_one = tree_one->next_v1;
}
else
{
i_buf--;
if( i_buf >= 0 )
{
CvPoint pt = ptr_buf[i_buf]->pt;
CV_WRITE_SEQ_ELEM( pt, writer );
tree_one = ptr_buf[i_buf]->next_v2;
if( log_iter )
{
cur_level = level_buf[i_buf] + 1;
}
}
}
}
contour = cvEndWriteSeq( &writer );
cvBoundingRect( contour, 1 );
__CLEANUP__;
__END__;
cvFree( &level_buf );
cvFree( &ptr_buf );
return contour;
}
CV_IMPL void
cvKMeans2( const CvArr* samples_arr, int cluster_count,
CvArr* labels_arr, CvTermCriteria termcrit )
{
CvMat* centers = 0;
CvMat* old_centers = 0;
CvMat* counters = 0;
CV_FUNCNAME( "cvKMeans2" );
__BEGIN__;
CvMat samples_stub, labels_stub;
CvMat* samples = (CvMat*)samples_arr;
CvMat* labels = (CvMat*)labels_arr;
CvMat* temp = 0;
CvRNG rng = CvRNG(-1);
int i, j, k, sample_count, dims;
int ids_delta, iter;
double max_dist;
if( !CV_IS_MAT( samples ))
CV_CALL( samples = cvGetMat( samples, &samples_stub ));
if( !CV_IS_MAT( labels ))
CV_CALL( labels = cvGetMat( labels, &labels_stub ));
if( cluster_count < 1 )
CV_ERROR( CV_StsOutOfRange, "Number of clusters should be positive" );
if( CV_MAT_DEPTH(samples->type) != CV_32F || CV_MAT_TYPE(labels->type) != CV_32SC1 )
CV_ERROR( CV_StsUnsupportedFormat,
"samples should be floating-point matrix, cluster_idx - integer vector" );
if( labels->rows != 1 && (labels->cols != 1 || !CV_IS_MAT_CONT(labels->type)) ||
labels->rows + labels->cols - 1 != samples->rows )
CV_ERROR( CV_StsUnmatchedSizes,
"cluster_idx should be 1D vector of the same number of elements as samples' number of rows" );
CV_CALL( termcrit = cvCheckTermCriteria( termcrit, 1e-6, 100 ));
termcrit.epsilon *= termcrit.epsilon;
sample_count = samples->rows;
if( cluster_count > sample_count )
cluster_count = sample_count;
dims = samples->cols*CV_MAT_CN(samples->type);
ids_delta = labels->step ? labels->step/(int)sizeof(int) : 1;
CV_CALL( centers = cvCreateMat( cluster_count, dims, CV_64FC1 ));
CV_CALL( old_centers = cvCreateMat( cluster_count, dims, CV_64FC1 ));
CV_CALL( counters = cvCreateMat( 1, cluster_count, CV_32SC1 ));
// init centers
for( i = 0; i < sample_count; i++ )
labels->data.i[i] = cvRandInt(&rng) % cluster_count;
counters->cols = cluster_count; // cut down counters
max_dist = termcrit.epsilon*2;
for( iter = 0; iter < termcrit.max_iter; iter++ )
{
// computer centers
cvZero( centers );
cvZero( counters );
for( i = 0; i < sample_count; i++ )
{
float* s = (float*)(samples->data.ptr + i*samples->step);
k = labels->data.i[i*ids_delta];
double* c = (double*)(centers->data.ptr + k*centers->step);
for( j = 0; j <= dims - 4; j += 4 )
{
double t0 = c[j] + s[j];
double t1 = c[j+1] + s[j+1];
c[j] = t0;
c[j+1] = t1;
t0 = c[j+2] + s[j+2];
t1 = c[j+3] + s[j+3];
c[j+2] = t0;
c[j+3] = t1;
}
for( ; j < dims; j++ )
c[j] += s[j];
counters->data.i[k]++;
}
if( iter > 0 )
max_dist = 0;
for( k = 0; k < cluster_count; k++ )
{
double* c = (double*)(centers->data.ptr + k*centers->step);
if( counters->data.i[k] != 0 )
{
double scale = 1./counters->data.i[k];
for( j = 0; j < dims; j++ )
c[j] *= scale;
}
else
{
i = cvRandInt( &rng ) % sample_count;
float* s = (float*)(samples->data.ptr + i*samples->step);
for( j = 0; j < dims; j++ )
c[j] = s[j];
}
if( iter > 0 )
{
double dist = 0;
double* c_o = (double*)(old_centers->data.ptr + k*old_centers->step);
for( j = 0; j < dims; j++ )
{
double t = c[j] - c_o[j];
dist += t*t;
}
if( max_dist < dist )
max_dist = dist;
}
}
// assign labels
for( i = 0; i < sample_count; i++ )
{
float* s = (float*)(samples->data.ptr + i*samples->step);
int k_best = 0;
double min_dist = DBL_MAX;
for( k = 0; k < cluster_count; k++ )
{
double* c = (double*)(centers->data.ptr + k*centers->step);
double dist = 0;
j = 0;
for( ; j <= dims - 4; j += 4 )
{
double t0 = c[j] - s[j];
double t1 = c[j+1] - s[j+1];
dist += t0*t0 + t1*t1;
t0 = c[j+2] - s[j+2];
t1 = c[j+3] - s[j+3];
dist += t0*t0 + t1*t1;
}
for( ; j < dims; j++ )
{
double t = c[j] - s[j];
dist += t*t;
}
if( min_dist > dist )
{
min_dist = dist;
k_best = k;
}
}
labels->data.i[i*ids_delta] = k_best;
}
if( max_dist < termcrit.epsilon )
break;
CV_SWAP( centers, old_centers, temp );
}
cvZero( counters );
for( i = 0; i < sample_count; i++ )
counters->data.i[labels->data.i[i]]++;
// ensure that we do not have empty clusters
for( k = 0; k < cluster_count; k++ )
if( counters->data.i[k] == 0 )
for(;;)
{
i = cvRandInt(&rng) % sample_count;
j = labels->data.i[i];
if( counters->data.i[j] > 1 )
{
labels->data.i[i] = k;
counters->data.i[j]--;
counters->data.i[k]++;
break;
}
}
__END__;
cvReleaseMat( ¢ers );
cvReleaseMat( &old_centers );
cvReleaseMat( &counters );
}
