opencv/modules/legacy/src/hmm.cpp

1699 lines
51 KiB
C++

/*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.
//
//
// Intel License Agreement
// For Open Source Computer Vision Library
//
// Copyright (C) 2000, Intel Corporation, 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 materials provided with the distribution.
//
// * The name of Intel Corporation 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 implied warranties, including, but not limited to, the implied
// 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"
#define LN2PI 1.837877f
#define BIG_FLT 1.e+10f
#define _CV_ERGODIC 1
#define _CV_CAUSAL 2
#define _CV_LAST_STATE 1
#define _CV_BEST_STATE 2
//*F///////////////////////////////////////////////////////////////////////////////////////
// Name: _cvCreateObsInfo
// Purpose: The function allocates memory for CvImgObsInfo structure
// and its inner stuff
// Context:
// Parameters: obs_info - addres of pointer to CvImgObsInfo structure
// num_hor_obs - number of horizontal observation vectors
// num_ver_obs - number of horizontal observation vectors
// obs_size - length of observation vector
//
// Returns: error status
//
// Notes:
//F*/
static CvStatus CV_STDCALL icvCreateObsInfo( CvImgObsInfo** obs_info,
CvSize num_obs, int obs_size )
{
int total = num_obs.height * num_obs.width;
CvImgObsInfo* obs = (CvImgObsInfo*)cvAlloc( sizeof( CvImgObsInfo) );
obs->obs_x = num_obs.width;
obs->obs_y = num_obs.height;
obs->obs = (float*)cvAlloc( total * obs_size * sizeof(float) );
obs->state = (int*)cvAlloc( 2 * total * sizeof(int) );
obs->mix = (int*)cvAlloc( total * sizeof(int) );
obs->obs_size = obs_size;
obs_info[0] = obs;
return CV_NO_ERR;
}
static CvStatus CV_STDCALL icvReleaseObsInfo( CvImgObsInfo** p_obs_info )
{
CvImgObsInfo* obs_info = p_obs_info[0];
cvFree( &(obs_info->obs) );
cvFree( &(obs_info->mix) );
cvFree( &(obs_info->state) );
cvFree( &(obs_info) );
p_obs_info[0] = NULL;
return CV_NO_ERR;
}
//*F///////////////////////////////////////////////////////////////////////////////////////
// Name: icvCreate2DHMM
// Purpose: The function allocates memory for 2-dimensional embedded HMM model
// and its inner stuff
// Context:
// Parameters: hmm - addres of pointer to CvEHMM structure
// state_number - array of hmm sizes (size of array == state_number[0]+1 )
// num_mix - number of gaussian mixtures in low-level HMM states
// size of array is defined by previous array values
// obs_size - length of observation vectors
//
// Returns: error status
//
// Notes: state_number[0] - number of states in external HMM.
// state_number[i] - number of states in embedded HMM
//
// example for face recognition: state_number = { 5 3 6 6 6 3 },
// length of num_mix array = 3+6+6+6+3 = 24//
//
//F*/
static CvStatus CV_STDCALL icvCreate2DHMM( CvEHMM** this_hmm,
int* state_number, int* num_mix, int obs_size )
{
int i;
int real_states = 0;
CvEHMMState* all_states;
CvEHMM* hmm;
int total_mix = 0;
float* pointers;
//compute total number of states of all level in 2d EHMM
for( i = 1; i <= state_number[0]; i++ )
{
real_states += state_number[i];
}
/* allocate memory for all hmms (from all levels) */
hmm = (CvEHMM*)cvAlloc( (state_number[0] + 1) * sizeof(CvEHMM) );
/* set number of superstates */
hmm[0].num_states = state_number[0];
hmm[0].level = 1;
/* allocate memory for all states */
all_states = (CvEHMMState *)cvAlloc( real_states * sizeof( CvEHMMState ) );
/* assign number of mixtures */
for( i = 0; i < real_states; i++ )
{
all_states[i].num_mix = num_mix[i];
}
/* compute size of inner of all real states */
for( i = 0; i < real_states; i++ )
{
total_mix += num_mix[i];
}
/* allocate memory for states stuff */
pointers = (float*)cvAlloc( total_mix * (2/*for mu invvar */ * obs_size +
2/*for weight and log_var_val*/ ) * sizeof( float) );
/* organize memory */
for( i = 0; i < real_states; i++ )
{
all_states[i].mu = pointers; pointers += num_mix[i] * obs_size;
all_states[i].inv_var = pointers; pointers += num_mix[i] * obs_size;
all_states[i].log_var_val = pointers; pointers += num_mix[i];
all_states[i].weight = pointers; pointers += num_mix[i];
}
/* set pointer to embedded hmm array */
hmm->u.ehmm = hmm + 1;
for( i = 0; i < hmm[0].num_states; i++ )
{
hmm[i+1].u.state = all_states;
all_states += state_number[i+1];
hmm[i+1].num_states = state_number[i+1];
}
for( i = 0; i <= state_number[0]; i++ )
{
hmm[i].transP = icvCreateMatrix_32f( hmm[i].num_states, hmm[i].num_states );
hmm[i].obsProb = NULL;
hmm[i].level = i ? 0 : 1;
}
/* if all ok - return pointer */
*this_hmm = hmm;
return CV_NO_ERR;
}
static CvStatus CV_STDCALL icvRelease2DHMM( CvEHMM** phmm )
{
CvEHMM* hmm = phmm[0];
int i;
for( i = 0; i < hmm[0].num_states + 1; i++ )
{
icvDeleteMatrix( hmm[i].transP );
}
if (hmm->obsProb != NULL)
{
int* tmp = ((int*)(hmm->obsProb)) - 3;
cvFree( &(tmp) );
}
cvFree( &(hmm->u.ehmm->u.state->mu) );
cvFree( &(hmm->u.ehmm->u.state) );
/* free hmm structures */
cvFree( phmm );
phmm[0] = NULL;
return CV_NO_ERR;
}
/* distance between 2 vectors */
static float icvSquareDistance( CvVect32f v1, CvVect32f v2, int len )
{
int i;
double dist0 = 0;
double dist1 = 0;
for( i = 0; i <= len - 4; i += 4 )
{
double t0 = v1[i] - v2[i];
double t1 = v1[i+1] - v2[i+1];
dist0 += t0*t0;
dist1 += t1*t1;
t0 = v1[i+2] - v2[i+2];
t1 = v1[i+3] - v2[i+3];
dist0 += t0*t0;
dist1 += t1*t1;
}
for( ; i < len; i++ )
{
double t0 = v1[i] - v2[i];
dist0 += t0*t0;
}
return (float)(dist0 + dist1);
}
/*can be used in CHMM & DHMM */
static CvStatus CV_STDCALL
icvUniformImgSegm( CvImgObsInfo* obs_info, CvEHMM* hmm )
{
#if 1
/* implementation is very bad */
int i, j, counter = 0;
CvEHMMState* first_state;
float inv_x = 1.f/obs_info->obs_x;
float inv_y = 1.f/obs_info->obs_y;
/* check arguments */
if ( !obs_info || !hmm ) return CV_NULLPTR_ERR;
first_state = hmm->u.ehmm->u.state;
for (i = 0; i < obs_info->obs_y; i++)
{
//bad line (division )
int superstate = (int)((i * hmm->num_states)*inv_y);/* /obs_info->obs_y; */
int index = (int)(hmm->u.ehmm[superstate].u.state - first_state);
for (j = 0; j < obs_info->obs_x; j++, counter++)
{
int state = (int)((j * hmm->u.ehmm[superstate].num_states)* inv_x); /* / obs_info->obs_x; */
obs_info->state[2 * counter] = superstate;
obs_info->state[2 * counter + 1] = state + index;
}
}
#else
//this is not ready yet
int i,j,k,m;
CvEHMMState* first_state = hmm->u.ehmm->u.state;
/* check bad arguments */
if ( hmm->num_states > obs_info->obs_y ) return CV_BADSIZE_ERR;
//compute vertical subdivision
float row_per_state = (float)obs_info->obs_y / hmm->num_states;
float col_per_state[1024]; /* maximum 1024 superstates */
//for every horizontal band compute subdivision
for( i = 0; i < hmm->num_states; i++ )
{
CvEHMM* ehmm = &(hmm->u.ehmm[i]);
col_per_state[i] = (float)obs_info->obs_x / ehmm->num_states;
}
//compute state bounds
int ss_bound[1024];
for( i = 0; i < hmm->num_states - 1; i++ )
{
ss_bound[i] = floor( row_per_state * ( i+1 ) );
}
ss_bound[hmm->num_states - 1] = obs_info->obs_y;
//work inside every superstate
int row = 0;
for( i = 0; i < hmm->num_states; i++ )
{
CvEHMM* ehmm = &(hmm->u.ehmm[i]);
int index = ehmm->u.state - first_state;
//calc distribution in superstate
int es_bound[1024];
for( j = 0; j < ehmm->num_states - 1; j++ )
{
es_bound[j] = floor( col_per_state[i] * ( j+1 ) );
}
es_bound[ehmm->num_states - 1] = obs_info->obs_x;
//assign states to first row of superstate
int col = 0;
for( j = 0; j < ehmm->num_states; j++ )
{
for( k = col; k < es_bound[j]; k++, col++ )
{
obs_info->state[row * obs_info->obs_x + 2 * k] = i;
obs_info->state[row * obs_info->obs_x + 2 * k + 1] = j + index;
}
col = es_bound[j];
}
//copy the same to other rows of superstate
for( m = row; m < ss_bound[i]; m++ )
{
memcpy( &(obs_info->state[m * obs_info->obs_x * 2]),
&(obs_info->state[row * obs_info->obs_x * 2]), obs_info->obs_x * 2 * sizeof(int) );
}
row = ss_bound[i];
}
#endif
return CV_NO_ERR;
}
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: InitMixSegm
// Purpose: The function implements the mixture segmentation of the states of the
// embedded HMM
// Context: used with the Viterbi training of the embedded HMM
// Function uses K-Means algorithm for clustering
//
// Parameters: obs_info_array - array of pointers to image observations
// num_img - length of above array
// hmm - pointer to HMM structure
//
// Returns: error status
//
// Notes:
//F*/
static CvStatus CV_STDCALL
icvInitMixSegm( CvImgObsInfo** obs_info_array, int num_img, CvEHMM* hmm )
{
int k, i, j;
int* num_samples; /* number of observations in every state */
int* counter; /* array of counters for every state */
int** a_class; /* for every state - characteristic array */
CvVect32f** samples; /* for every state - pointer to observation vectors */
int*** samples_mix; /* for every state - array of pointers to vectors mixtures */
CvTermCriteria criteria = cvTermCriteria( CV_TERMCRIT_EPS|CV_TERMCRIT_ITER,
1000, /* iter */
0.01f ); /* eps */
int total = 0;
CvEHMMState* first_state = hmm->u.ehmm->u.state;
for( i = 0 ; i < hmm->num_states; i++ )
{
total += hmm->u.ehmm[i].num_states;
}
/* for every state integer is allocated - number of vectors in state */
num_samples = (int*)cvAlloc( total * sizeof(int) );
/* integer counter is allocated for every state */
counter = (int*)cvAlloc( total * sizeof(int) );
samples = (CvVect32f**)cvAlloc( total * sizeof(CvVect32f*) );
samples_mix = (int***)cvAlloc( total * sizeof(int**) );
/* clear */
memset( num_samples, 0 , total*sizeof(int) );
memset( counter, 0 , total*sizeof(int) );
/* for every state the number of vectors which belong to it is computed (smth. like histogram) */
for (k = 0; k < num_img; k++)
{
CvImgObsInfo* obs = obs_info_array[k];
int count = 0;
for (i = 0; i < obs->obs_y; i++)
{
for (j = 0; j < obs->obs_x; j++, count++)
{
int state = obs->state[ 2 * count + 1];
num_samples[state] += 1;
}
}
}
/* for every state int* is allocated */
a_class = (int**)cvAlloc( total*sizeof(int*) );
for (i = 0; i < total; i++)
{
a_class[i] = (int*)cvAlloc( num_samples[i] * sizeof(int) );
samples[i] = (CvVect32f*)cvAlloc( num_samples[i] * sizeof(CvVect32f) );
samples_mix[i] = (int**)cvAlloc( num_samples[i] * sizeof(int*) );
}
/* for every state vectors which belong to state are gathered */
for (k = 0; k < num_img; k++)
{
CvImgObsInfo* obs = obs_info_array[k];
int num_obs = ( obs->obs_x ) * ( obs->obs_y );
float* vector = obs->obs;
for (i = 0; i < num_obs; i++, vector+=obs->obs_size )
{
int state = obs->state[2*i+1];
samples[state][counter[state]] = vector;
samples_mix[state][counter[state]] = &(obs->mix[i]);
counter[state]++;
}
}
/* clear counters */
memset( counter, 0, total*sizeof(int) );
/* do the actual clustering using the K Means algorithm */
for (i = 0; i < total; i++)
{
if ( first_state[i].num_mix == 1)
{
for (k = 0; k < num_samples[i]; k++)
{
/* all vectors belong to one mixture */
a_class[i][k] = 0;
}
}
else if( num_samples[i] )
{
/* clusterize vectors */
cvKMeans( first_state[i].num_mix, samples[i], num_samples[i],
obs_info_array[0]->obs_size, criteria, a_class[i] );
}
}
/* for every vector number of mixture is assigned */
for( i = 0; i < total; i++ )
{
for (j = 0; j < num_samples[i]; j++)
{
samples_mix[i][j][0] = a_class[i][j];
}
}
for (i = 0; i < total; i++)
{
cvFree( &(a_class[i]) );
cvFree( &(samples[i]) );
cvFree( &(samples_mix[i]) );
}
cvFree( &a_class );
cvFree( &samples );
cvFree( &samples_mix );
cvFree( &counter );
cvFree( &num_samples );
return CV_NO_ERR;
}
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: ComputeUniModeGauss
// Purpose: The function computes the Gaussian pdf for a sample vector
// Context:
// Parameters: obsVeq - pointer to the sample vector
// mu - pointer to the mean vector of the Gaussian pdf
// var - pointer to the variance vector of the Gaussian pdf
// VecSize - the size of sample vector
//
// Returns: the pdf of the sample vector given the specified Gaussian
//
// Notes:
//F*/
/*static float icvComputeUniModeGauss(CvVect32f vect, CvVect32f mu,
CvVect32f inv_var, float log_var_val, int vect_size)
{
int n;
double tmp;
double prob;
prob = -log_var_val;
for (n = 0; n < vect_size; n++)
{
tmp = (vect[n] - mu[n]) * inv_var[n];
prob = prob - tmp * tmp;
}
//prob *= 0.5f;
return (float)prob;
}*/
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: ComputeGaussMixture
// Purpose: The function computes the mixture Gaussian pdf of a sample vector.
// Context:
// Parameters: obsVeq - pointer to the sample vector
// mu - two-dimensional pointer to the mean vector of the Gaussian pdf;
// the first dimension is indexed over the number of mixtures and
// the second dimension is indexed along the size of the mean vector
// var - two-dimensional pointer to the variance vector of the Gaussian pdf;
// the first dimension is indexed over the number of mixtures and
// the second dimension is indexed along the size of the variance vector
// VecSize - the size of sample vector
// weight - pointer to the wights of the Gaussian mixture
// NumMix - the number of Gaussian mixtures
//
// Returns: the pdf of the sample vector given the specified Gaussian mixture.
//
// Notes:
//F*/
/* Calculate probability of observation at state in logarithmic scale*/
/*static float
icvComputeGaussMixture( CvVect32f vect, float* mu,
float* inv_var, float* log_var_val,
int vect_size, float* weight, int num_mix )
{
double prob, l_prob;
prob = 0.0f;
if (num_mix == 1)
{
return icvComputeUniModeGauss( vect, mu, inv_var, log_var_val[0], vect_size);
}
else
{
int m;
for (m = 0; m < num_mix; m++)
{
if ( weight[m] > 0.0)
{
l_prob = icvComputeUniModeGauss(vect, mu + m*vect_size,
inv_var + m * vect_size,
log_var_val[m],
vect_size);
prob = prob + weight[m]*exp((double)l_prob);
}
}
prob = log(prob);
}
return (float)prob;
}*/
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: EstimateObsProb
// Purpose: The function computes the probability of every observation in every state
// Context:
// Parameters: obs_info - observations
// hmm - hmm
// Returns: error status
//
// Notes:
//F*/
static CvStatus CV_STDCALL icvEstimateObsProb( CvImgObsInfo* obs_info, CvEHMM* hmm )
{
int i, j;
int total_states = 0;
/* check if matrix exist and check current size
if not sufficient - realloc */
int status = 0; /* 1 - not allocated, 2 - allocated but small size,
3 - size is enough, but distribution is bad, 0 - all ok */
for( j = 0; j < hmm->num_states; j++ )
{
total_states += hmm->u.ehmm[j].num_states;
}
if ( hmm->obsProb == NULL )
{
/* allocare memory */
int need_size = ( obs_info->obs_x * obs_info->obs_y * total_states * sizeof(float) +
obs_info->obs_y * hmm->num_states * sizeof( CvMatr32f) );
int* buffer = (int*)cvAlloc( need_size + 3 * sizeof(int) );
buffer[0] = need_size;
buffer[1] = obs_info->obs_y;
buffer[2] = obs_info->obs_x;
hmm->obsProb = (float**) (buffer + 3);
status = 3;
}
else
{
/* check current size */
int* total= (int*)(((int*)(hmm->obsProb)) - 3);
int need_size = ( obs_info->obs_x * obs_info->obs_y * total_states * sizeof(float) +
obs_info->obs_y * hmm->num_states * sizeof( CvMatr32f/*(float*)*/ ) );
assert( sizeof(float*) == sizeof(int) );
if ( need_size > (*total) )
{
int* buffer = ((int*)(hmm->obsProb)) - 3;
cvFree( &buffer);
buffer = (int*)cvAlloc( need_size + 3 * sizeof(int));
buffer[0] = need_size;
buffer[1] = obs_info->obs_y;
buffer[2] = obs_info->obs_x;
hmm->obsProb = (float**)(buffer + 3);
status = 3;
}
}
if (!status)
{
int* obsx = ((int*)(hmm->obsProb)) - 1;
int* obsy = ((int*)(hmm->obsProb)) - 2;
assert( (*obsx > 0) && (*obsy > 0) );
/* is good distribution? */
if ( (obs_info->obs_x > (*obsx) ) || (obs_info->obs_y > (*obsy) ) )
status = 3;
}
/* if bad status - do reallocation actions */
assert( (status == 0) || (status == 3) );
if ( status )
{
float** tmp = hmm->obsProb;
float* tmpf;
/* distribute pointers of ehmm->obsProb */
for( i = 0; i < hmm->num_states; i++ )
{
hmm->u.ehmm[i].obsProb = tmp;
tmp += obs_info->obs_y;
}
tmpf = (float*)tmp;
/* distribute pointers of ehmm->obsProb[j] */
for( i = 0; i < hmm->num_states; i++ )
{
CvEHMM* ehmm = &( hmm->u.ehmm[i] );
for( j = 0; j < obs_info->obs_y; j++ )
{
ehmm->obsProb[j] = tmpf;
tmpf += ehmm->num_states * obs_info->obs_x;
}
}
}/* end of pointer distribution */
#if 1
{
#define MAX_BUF_SIZE 1200
float local_log_mix_prob[MAX_BUF_SIZE];
double local_mix_prob[MAX_BUF_SIZE];
int vect_size = obs_info->obs_size;
CvStatus res = CV_NO_ERR;
float* log_mix_prob = local_log_mix_prob;
double* mix_prob = local_mix_prob;
int max_size = 0;
int obs_x = obs_info->obs_x;
/* calculate temporary buffer size */
for( i = 0; i < hmm->num_states; i++ )
{
CvEHMM* ehmm = &(hmm->u.ehmm[i]);
CvEHMMState* state = ehmm->u.state;
int max_mix = 0;
for( j = 0; j < ehmm->num_states; j++ )
{
int t = state[j].num_mix;
if( max_mix < t ) max_mix = t;
}
max_mix *= ehmm->num_states;
if( max_size < max_mix ) max_size = max_mix;
}
max_size *= obs_x * vect_size;
/* allocate buffer */
if( max_size > MAX_BUF_SIZE )
{
log_mix_prob = (float*)cvAlloc( max_size*(sizeof(float) + sizeof(double)));
if( !log_mix_prob ) return CV_OUTOFMEM_ERR;
mix_prob = (double*)(log_mix_prob + max_size);
}
memset( log_mix_prob, 0, max_size*sizeof(float));
/*****************computing probabilities***********************/
/* loop through external states */
for( i = 0; i < hmm->num_states; i++ )
{
CvEHMM* ehmm = &(hmm->u.ehmm[i]);
CvEHMMState* state = ehmm->u.state;
int max_mix = 0;
int n_states = ehmm->num_states;
/* determine maximal number of mixtures (again) */
for( j = 0; j < ehmm->num_states; j++ )
{
int t = state[j].num_mix;
if( max_mix < t ) max_mix = t;
}
/* loop through rows of the observation matrix */
for( j = 0; j < obs_info->obs_y; j++ )
{
int m, n;
float* obs = obs_info->obs + j * obs_x * vect_size;
float* log_mp = max_mix > 1 ? log_mix_prob : ehmm->obsProb[j];
double* mp = mix_prob;
/* several passes are done below */
/* 1. calculate logarithms of probabilities for each mixture */
/* loop through mixtures */
for( m = 0; m < max_mix; m++ )
{
/* set pointer to first observation in the line */
float* vect = obs;
/* cycles through obseravtions in the line */
for( n = 0; n < obs_x; n++, vect += vect_size, log_mp += n_states )
{
int k, l;
for( l = 0; l < n_states; l++ )
{
if( state[l].num_mix > m )
{
float* mu = state[l].mu + m*vect_size;
float* inv_var = state[l].inv_var + m*vect_size;
double prob = -state[l].log_var_val[m];
for( k = 0; k < vect_size; k++ )
{
double t = (vect[k] - mu[k])*inv_var[k];
prob -= t*t;
}
log_mp[l] = MAX( (float)prob, -500 );
}
}
}
}
/* skip the rest if there is a single mixture */
if( max_mix == 1 ) continue;
/* 2. calculate exponent of log_mix_prob
(i.e. probability for each mixture) */
cvbFastExp( log_mix_prob, mix_prob, max_mix * obs_x * n_states );
/* 3. sum all mixtures with weights */
/* 3a. first mixture - simply scale by weight */
for( n = 0; n < obs_x; n++, mp += n_states )
{
int l;
for( l = 0; l < n_states; l++ )
{
mp[l] *= state[l].weight[0];
}
}
/* 3b. add other mixtures */
for( m = 1; m < max_mix; m++ )
{
int ofs = -m*obs_x*n_states;
for( n = 0; n < obs_x; n++, mp += n_states )
{
int l;
for( l = 0; l < n_states; l++ )
{
if( m < state[l].num_mix )
{
mp[l + ofs] += mp[l] * state[l].weight[m];
}
}
}
}
/* 4. Put logarithms of summary probabilities to the destination matrix */
cvbFastLog( mix_prob, ehmm->obsProb[j], obs_x * n_states );
}
}
if( log_mix_prob != local_log_mix_prob ) cvFree( &log_mix_prob );
return res;
#undef MAX_BUF_SIZE
}
#else
for( i = 0; i < hmm->num_states; i++ )
{
CvEHMM* ehmm = &(hmm->u.ehmm[i]);
CvEHMMState* state = ehmm->u.state;
for( j = 0; j < obs_info->obs_y; j++ )
{
int k,m;
int obs_index = j * obs_info->obs_x;
float* B = ehmm->obsProb[j];
/* cycles through obs and states */
for( k = 0; k < obs_info->obs_x; k++ )
{
CvVect32f vect = (obs_info->obs) + (obs_index + k) * vect_size;
float* matr_line = B + k * ehmm->num_states;
for( m = 0; m < ehmm->num_states; m++ )
{
matr_line[m] = icvComputeGaussMixture( vect, state[m].mu, state[m].inv_var,
state[m].log_var_val, vect_size, state[m].weight,
state[m].num_mix );
}
}
}
}
#endif
}
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: EstimateTransProb
// Purpose: The function calculates the state and super state transition probabilities
// of the model given the images,
// the state segmentation and the input parameters
// Context:
// Parameters: obs_info_array - array of pointers to image observations
// num_img - length of above array
// hmm - pointer to HMM structure
// Returns: void
//
// Notes:
//F*/
static CvStatus CV_STDCALL
icvEstimateTransProb( CvImgObsInfo** obs_info_array, int num_img, CvEHMM* hmm )
{
int i, j, k;
CvEHMMState* first_state = hmm->u.ehmm->u.state;
/* as a counter we will use transP matrix */
/* initialization */
/* clear transP */
icvSetZero_32f( hmm->transP, hmm->num_states, hmm->num_states );
for (i = 0; i < hmm->num_states; i++ )
{
icvSetZero_32f( hmm->u.ehmm[i].transP , hmm->u.ehmm[i].num_states, hmm->u.ehmm[i].num_states );
}
/* compute the counters */
for (i = 0; i < num_img; i++)
{
int counter = 0;
CvImgObsInfo* info = obs_info_array[i];
for (j = 0; j < info->obs_y; j++)
{
for (k = 0; k < info->obs_x; k++, counter++)
{
/* compute how many transitions from state to state
occured both in horizontal and vertical direction */
int superstate, state;
int nextsuperstate, nextstate;
int begin_ind;
superstate = info->state[2 * counter];
begin_ind = (int)(hmm->u.ehmm[superstate].u.state - first_state);
state = info->state[ 2 * counter + 1] - begin_ind;
if (j < info->obs_y - 1)
{
int transP_size = hmm->num_states;
nextsuperstate = info->state[ 2*(counter + info->obs_x) ];
hmm->transP[superstate * transP_size + nextsuperstate] += 1;
}
if (k < info->obs_x - 1)
{
int transP_size = hmm->u.ehmm[superstate].num_states;
nextstate = info->state[2*(counter+1) + 1] - begin_ind;
hmm->u.ehmm[superstate].transP[ state * transP_size + nextstate] += 1;
}
}
}
}
/* estimate superstate matrix */
for( i = 0; i < hmm->num_states; i++)
{
float total = 0;
float inv_total;
for( j = 0; j < hmm->num_states; j++)
{
total += hmm->transP[i * hmm->num_states + j];
}
//assert( total );
inv_total = total ? 1.f/total : 0;
for( j = 0; j < hmm->num_states; j++)
{
hmm->transP[i * hmm->num_states + j] =
hmm->transP[i * hmm->num_states + j] ?
(float)log( hmm->transP[i * hmm->num_states + j] * inv_total ) : -BIG_FLT;
}
}
/* estimate other matrices */
for( k = 0; k < hmm->num_states; k++ )
{
CvEHMM* ehmm = &(hmm->u.ehmm[k]);
for( i = 0; i < ehmm->num_states; i++)
{
float total = 0;
float inv_total;
for( j = 0; j < ehmm->num_states; j++)
{
total += ehmm->transP[i*ehmm->num_states + j];
}
//assert( total );
inv_total = total ? 1.f/total : 0;
for( j = 0; j < ehmm->num_states; j++)
{
ehmm->transP[i * ehmm->num_states + j] =
(ehmm->transP[i * ehmm->num_states + j]) ?
(float)log( ehmm->transP[i * ehmm->num_states + j] * inv_total) : -BIG_FLT ;
}
}
}
return CV_NO_ERR;
}
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: MixSegmL2
// Purpose: The function implements the mixture segmentation of the states of the
// embedded HMM
// Context: used with the Viterbi training of the embedded HMM
//
// Parameters:
// obs_info_array
// num_img
// hmm
// Returns: void
//
// Notes:
//F*/
static CvStatus CV_STDCALL
icvMixSegmL2( CvImgObsInfo** obs_info_array, int num_img, CvEHMM* hmm )
{
int k, i, j, m;
CvEHMMState* state = hmm->u.ehmm[0].u.state;
for (k = 0; k < num_img; k++)
{
int counter = 0;
CvImgObsInfo* info = obs_info_array[k];
for (i = 0; i < info->obs_y; i++)
{
for (j = 0; j < info->obs_x; j++, counter++)
{
int e_state = info->state[2 * counter + 1];
float min_dist;
min_dist = icvSquareDistance((info->obs) + (counter * info->obs_size),
state[e_state].mu, info->obs_size);
info->mix[counter] = 0;
for (m = 1; m < state[e_state].num_mix; m++)
{
float dist=icvSquareDistance( (info->obs) + (counter * info->obs_size),
state[e_state].mu + m * info->obs_size,
info->obs_size);
if (dist < min_dist)
{
min_dist = dist;
/* assign mixture with smallest distance */
info->mix[counter] = m;
}
}
}
}
}
return CV_NO_ERR;
}
/*
CvStatus icvMixSegmProb(CvImgObsInfo* obs_info, int num_img, CvEHMM* hmm )
{
int k, i, j, m;
CvEHMMState* state = hmm->ehmm[0].state_info;
for (k = 0; k < num_img; k++)
{
int counter = 0;
CvImgObsInfo* info = obs_info + k;
for (i = 0; i < info->obs_y; i++)
{
for (j = 0; j < info->obs_x; j++, counter++)
{
int e_state = info->in_state[counter];
float max_prob;
max_prob = icvComputeUniModeGauss( info->obs[counter], state[e_state].mu[0],
state[e_state].inv_var[0],
state[e_state].log_var[0],
info->obs_size );
info->mix[counter] = 0;
for (m = 1; m < state[e_state].num_mix; m++)
{
float prob=icvComputeUniModeGauss(info->obs[counter], state[e_state].mu[m],
state[e_state].inv_var[m],
state[e_state].log_var[m],
info->obs_size);
if (prob > max_prob)
{
max_prob = prob;
// assign mixture with greatest probability.
info->mix[counter] = m;
}
}
}
}
}
return CV_NO_ERR;
}
*/
static CvStatus CV_STDCALL
icvViterbiSegmentation( int num_states, int /*num_obs*/, CvMatr32f transP,
CvMatr32f B, int start_obs, int prob_type,
int** q, int min_num_obs, int max_num_obs,
float* prob )
{
// memory allocation
int i, j, last_obs;
int m_HMMType = _CV_ERGODIC; /* _CV_CAUSAL or _CV_ERGODIC */
int m_ProbType = prob_type; /* _CV_LAST_STATE or _CV_BEST_STATE */
int m_minNumObs = min_num_obs; /*??*/
int m_maxNumObs = max_num_obs; /*??*/
int m_numStates = num_states;
float* m_pi = (float*)cvAlloc( num_states* sizeof(float) );
CvMatr32f m_a = transP;
// offset brobability matrix to starting observation
CvMatr32f m_b = B + start_obs * num_states;
//so m_xl will not be used more
//m_xl = start_obs;
/* if (muDur != NULL){
m_d = new int[m_numStates];
m_l = new double[m_numStates];
for (i = 0; i < m_numStates; i++){
m_l[i] = muDur[i];
}
}
else{
m_d = NULL;
m_l = NULL;
}
*/
CvMatr32f m_Gamma = icvCreateMatrix_32f( num_states, m_maxNumObs );
int* m_csi = (int*)cvAlloc( num_states * m_maxNumObs * sizeof(int) );
//stores maximal result for every ending observation */
CvVect32f m_MaxGamma = prob;
// assert( m_xl + max_num_obs <= num_obs );
/*??m_q = new int*[m_maxNumObs - m_minNumObs];
??for (i = 0; i < m_maxNumObs - m_minNumObs; i++)
?? m_q[i] = new int[m_minNumObs + i + 1];
*/
/******************************************************************/
/* Viterbi initialization */
/* set initial state probabilities, in logarithmic scale */
for (i = 0; i < m_numStates; i++)
{
m_pi[i] = -BIG_FLT;
}
m_pi[0] = 0.0f;
for (i = 0; i < num_states; i++)
{
m_Gamma[0 * num_states + i] = m_pi[i] + m_b[0 * num_states + i];
m_csi[0 * num_states + i] = 0;
}
/******************************************************************/
/* Viterbi recursion */
if ( m_HMMType == _CV_CAUSAL ) //causal model
{
int t;
for (t = 1 ; t < m_maxNumObs; t++)
{
// evaluate self-to-self transition for state 0
m_Gamma[t * num_states + 0] = m_Gamma[(t-1) * num_states + 0] + m_a[0];
m_csi[t * num_states + 0] = 0;
for (j = 1; j < num_states; j++)
{
float self = m_Gamma[ (t-1) * num_states + j] + m_a[ j * num_states + j];
float prev = m_Gamma[ (t-1) * num_states +(j-1)] + m_a[ (j-1) * num_states + j];
if ( prev > self )
{
m_csi[t * num_states + j] = j-1;
m_Gamma[t * num_states + j] = prev;
}
else
{
m_csi[t * num_states + j] = j;
m_Gamma[t * num_states + j] = self;
}
m_Gamma[t * num_states + j] = m_Gamma[t * num_states + j] + m_b[t * num_states + j];
}
}
}
else if ( m_HMMType == _CV_ERGODIC ) //ergodic model
{
int t;
for (t = 1 ; t < m_maxNumObs; t++)
{
for (j = 0; j < num_states; j++)
{
m_Gamma[ t*num_states + j] = m_Gamma[(t-1) * num_states + 0] + m_a[0*num_states+j];
m_csi[t *num_states + j] = 0;
for (i = 1; i < num_states; i++)
{
float currGamma = m_Gamma[(t-1) *num_states + i] + m_a[i *num_states + j];
if (currGamma > m_Gamma[t *num_states + j])
{
m_Gamma[t * num_states + j] = currGamma;
m_csi[t * num_states + j] = i;
}
}
m_Gamma[t *num_states + j] = m_Gamma[t *num_states + j] + m_b[t * num_states + j];
}
}
}
for( last_obs = m_minNumObs-1, i = 0; last_obs < m_maxNumObs; last_obs++, i++ )
{
int t;
/******************************************************************/
/* Viterbi termination */
if ( m_ProbType == _CV_LAST_STATE )
{
m_MaxGamma[i] = m_Gamma[last_obs * num_states + num_states - 1];
q[i][last_obs] = num_states - 1;
}
else if( m_ProbType == _CV_BEST_STATE )
{
int k;
q[i][last_obs] = 0;
m_MaxGamma[i] = m_Gamma[last_obs * num_states + 0];
for(k = 1; k < num_states; k++)
{
if ( m_Gamma[last_obs * num_states + k] > m_MaxGamma[i] )
{
m_MaxGamma[i] = m_Gamma[last_obs * num_states + k];
q[i][last_obs] = k;
}
}
}
/******************************************************************/
/* Viterbi backtracking */
for (t = last_obs-1; t >= 0; t--)
{
q[i][t] = m_csi[(t+1) * num_states + q[i][t+1] ];
}
}
/* memory free */
cvFree( &m_pi );
cvFree( &m_csi );
icvDeleteMatrix( m_Gamma );
return CV_NO_ERR;
}
/*F///////////////////////////////////////////////////////////////////////////////////////
// Name: icvEViterbi
// Purpose: The function calculates the embedded Viterbi algorithm
// for 1 image
// Context:
// Parameters:
// obs_info - observations
// hmm - HMM
//
// Returns: the Embedded Viterbi probability (float)
// and do state segmentation of observations
//
// Notes:
//F*/
static float CV_STDCALL icvEViterbi( CvImgObsInfo* obs_info, CvEHMM* hmm )
{
int i, j, counter;
float log_likelihood;
float inv_obs_x = 1.f / obs_info->obs_x;
CvEHMMState* first_state = hmm->u.ehmm->u.state;
/* memory allocation for superB */
CvMatr32f superB = icvCreateMatrix_32f(hmm->num_states, obs_info->obs_y );
/* memory allocation for q */
int*** q = (int***)cvAlloc( hmm->num_states * sizeof(int**) );
int* super_q = (int*)cvAlloc( obs_info->obs_y * sizeof(int) );
for (i = 0; i < hmm->num_states; i++)
{
q[i] = (int**)cvAlloc( obs_info->obs_y * sizeof(int*) );
for (j = 0; j < obs_info->obs_y ; j++)
{
q[i][j] = (int*)cvAlloc( obs_info->obs_x * sizeof(int) );
}
}
/* start Viterbi segmentation */
for (i = 0; i < hmm->num_states; i++)
{
CvEHMM* ehmm = &(hmm->u.ehmm[i]);
for (j = 0; j < obs_info->obs_y; j++)
{
float max_gamma;
/* 1D HMM Viterbi segmentation */
icvViterbiSegmentation( ehmm->num_states, obs_info->obs_x,
ehmm->transP, ehmm->obsProb[j], 0,
_CV_LAST_STATE, &q[i][j], obs_info->obs_x,
obs_info->obs_x, &max_gamma);
superB[j * hmm->num_states + i] = max_gamma * inv_obs_x;
}
}
/* perform global Viterbi segmentation (i.e. process higher-level HMM) */
icvViterbiSegmentation( hmm->num_states, obs_info->obs_y,
hmm->transP, superB, 0,
_CV_LAST_STATE, &super_q, obs_info->obs_y,
obs_info->obs_y, &log_likelihood );
log_likelihood /= obs_info->obs_y ;
counter = 0;
/* assign new state to observation vectors */
for (i = 0; i < obs_info->obs_y; i++)
{
for (j = 0; j < obs_info->obs_x; j++, counter++)
{
int superstate = super_q[i];
int state = (int)(hmm->u.ehmm[superstate].u.state - first_state);
obs_info->state[2 * counter] = superstate;
obs_info->state[2 * counter + 1] = state + q[superstate][i][j];
}
}
/* memory deallocation for superB */
icvDeleteMatrix( superB );
/*memory deallocation for q */
for (i = 0; i < hmm->num_states; i++)
{
for (j = 0; j < obs_info->obs_y ; j++)
{
cvFree( &q[i][j] );
}
cvFree( &q[i] );
}
cvFree( &q );
cvFree( &super_q );
return log_likelihood;
}
static CvStatus CV_STDCALL
icvEstimateHMMStateParams( CvImgObsInfo** obs_info_array, int num_img, CvEHMM* hmm )
{
/* compute gamma, weights, means, vars */
int k, i, j, m;
int total = 0;
int vect_len = obs_info_array[0]->obs_size;
float start_log_var_val = LN2PI * vect_len;
CvVect32f tmp_vect = icvCreateVector_32f( vect_len );
CvEHMMState* first_state = hmm->u.ehmm[0].u.state;
assert( sizeof(float) == sizeof(int) );
for(i = 0; i < hmm->num_states; i++ )
{
total+= hmm->u.ehmm[i].num_states;
}
/***************Gamma***********************/
/* initialize gamma */
for( i = 0; i < total; i++ )
{
for (m = 0; m < first_state[i].num_mix; m++)
{
((int*)(first_state[i].weight))[m] = 0;
}
}
/* maybe gamma must be computed in mixsegm process ?? */
/* compute gamma */
for (k = 0; k < num_img; k++)
{
CvImgObsInfo* info = obs_info_array[k];
int num_obs = info->obs_y * info->obs_x;
for (i = 0; i < num_obs; i++)
{
int state, mixture;
state = info->state[2*i + 1];
mixture = info->mix[i];
/* computes gamma - number of observations corresponding
to every mixture of every state */
((int*)(first_state[state].weight))[mixture] += 1;
}
}
/***************Mean and Var***********************/
/* compute means and variances of every item */
/* initially variance placed to inv_var */
/* zero mean and variance */
for (i = 0; i < total; i++)
{
memset( (void*)first_state[i].mu, 0, first_state[i].num_mix * vect_len *
sizeof(float) );
memset( (void*)first_state[i].inv_var, 0, first_state[i].num_mix * vect_len *
sizeof(float) );
}
/* compute sums */
for (i = 0; i < num_img; i++)
{
CvImgObsInfo* info = obs_info_array[i];
int total_obs = info->obs_x * info->obs_y;
float* vector = info->obs;
for (j = 0; j < total_obs; j++, vector+=vect_len )
{
int state = info->state[2 * j + 1];
int mixture = info->mix[j];
CvVect32f mean = first_state[state].mu + mixture * vect_len;
CvVect32f mean2 = first_state[state].inv_var + mixture * vect_len;
icvAddVector_32f( mean, vector, mean, vect_len );
for( k = 0; k < vect_len; k++ )
mean2[k] += vector[k]*vector[k];
}
}
/*compute the means and variances */
/* assume gamma already computed */
for (i = 0; i < total; i++)
{
CvEHMMState* state = &(first_state[i]);
for (m = 0; m < state->num_mix; m++)
{
CvVect32f mu = state->mu + m * vect_len;
CvVect32f invar = state->inv_var + m * vect_len;
if ( ((int*)state->weight)[m] > 1)
{
float inv_gamma = 1.f/((int*)(state->weight))[m];
icvScaleVector_32f( mu, mu, vect_len, inv_gamma);
icvScaleVector_32f( invar, invar, vect_len, inv_gamma);
}
icvMulVectors_32f(mu, mu, tmp_vect, vect_len);
icvSubVector_32f( invar, tmp_vect, invar, vect_len);
/* low bound of variance - 100 (Ara's experimental result) */
for( k = 0; k < vect_len; k++ )
{
invar[k] = (invar[k] > 100.f) ? invar[k] : 100.f;
}
/* compute log_var */
state->log_var_val[m] = start_log_var_val;
for( k = 0; k < vect_len; k++ )
{
state->log_var_val[m] += (float)log( invar[k] );
}
/* SMOLI 27.10.2000 */
state->log_var_val[m] *= 0.5;
/* compute inv_var = 1/sqrt(2*variance) */
icvScaleVector_32f(invar, invar, vect_len, 2.f );
cvbInvSqrt( invar, invar, vect_len );
}
}
/***************Weights***********************/
/* normilize gammas - i.e. compute mixture weights */
//compute weights
for (i = 0; i < total; i++)
{
int gamma_total = 0;
float norm;
for (m = 0; m < first_state[i].num_mix; m++)
{
gamma_total += ((int*)(first_state[i].weight))[m];
}
norm = gamma_total ? (1.f/(float)gamma_total) : 0.f;
for (m = 0; m < first_state[i].num_mix; m++)
{
first_state[i].weight[m] = ((int*)(first_state[i].weight))[m] * norm;
}
}
icvDeleteVector( tmp_vect);
return CV_NO_ERR;
}
/*
CvStatus icvLightingCorrection8uC1R( uchar* img, CvSize roi, int src_step )
{
int i, j;
int width = roi.width;
int height = roi.height;
float x1, x2, y1, y2;
int f[3] = {0, 0, 0};
float a[3] = {0, 0, 0};
float h1;
float h2;
float c1,c2;
float min = FLT_MAX;
float max = -FLT_MAX;
float correction;
float* float_img = icvAlloc( width * height * sizeof(float) );
x1 = width * (width + 1) / 2.0f; // Sum (1, ... , width)
x2 = width * (width + 1 ) * (2 * width + 1) / 6.0f; // Sum (1^2, ... , width^2)
y1 = height * (height + 1)/2.0f; // Sum (1, ... , width)
y2 = height * (height + 1 ) * (2 * height + 1) / 6.0f; // Sum (1^2, ... , width^2)
// extract grayvalues
for (i = 0; i < height; i++)
{
for (j = 0; j < width; j++)
{
f[2] = f[2] + j * img[i*src_step + j];
f[1] = f[1] + i * img[i*src_step + j];
f[0] = f[0] + img[i*src_step + j];
}
}
h1 = (float)f[0] * (float)x1 / (float)width;
h2 = (float)f[0] * (float)y1 / (float)height;
a[2] = ((float)f[2] - h1) / (float)(x2*height - x1*x1*height/(float)width);
a[1] = ((float)f[1] - h2) / (float)(y2*width - y1*y1*width/(float)height);
a[0] = (float)f[0]/(float)(width*height) - (float)y1*a[1]/(float)height -
(float)x1*a[2]/(float)width;
for (i = 0; i < height; i++)
{
for (j = 0; j < width; j++)
{
correction = a[0] + a[1]*(float)i + a[2]*(float)j;
float_img[i*width + j] = img[i*src_step + j] - correction;
if (float_img[i*width + j] < min) min = float_img[i*width+j];
if (float_img[i*width + j] > max) max = float_img[i*width+j];
}
}
//rescaling to the range 0:255
c2 = 0;
if (max == min)
c2 = 255.0f;
else
c2 = 255.0f/(float)(max - min);
c1 = (-(float)min)*c2;
for (i = 0; i < height; i++)
{
for (j = 0; j < width; j++)
{
int value = (int)floor(c2*float_img[i*width + j] + c1);
if (value < 0) value = 0;
if (value > 255) value = 255;
img[i*src_step + j] = (uchar)value;
}
}
cvFree( &float_img );
return CV_NO_ERR;
}
CvStatus icvLightingCorrection( icvImage* img )
{
CvSize roi;
if ( img->type != IPL_DEPTH_8U || img->channels != 1 )
return CV_BADFACTOR_ERR;
roi = _cvSize( img->roi.width, img->roi.height );
return _cvLightingCorrection8uC1R( img->data + img->roi.y * img->step + img->roi.x,
roi, img->step );
}
*/
CV_IMPL CvEHMM*
cvCreate2DHMM( int *state_number, int *num_mix, int obs_size )
{
CvEHMM* hmm = 0;
IPPI_CALL( icvCreate2DHMM( &hmm, state_number, num_mix, obs_size ));
return hmm;
}
CV_IMPL void
cvRelease2DHMM( CvEHMM ** hmm )
{
IPPI_CALL( icvRelease2DHMM( hmm ));
}
CV_IMPL CvImgObsInfo*
cvCreateObsInfo( CvSize num_obs, int obs_size )
{
CvImgObsInfo *obs_info = 0;
IPPI_CALL( icvCreateObsInfo( &obs_info, num_obs, obs_size ));
return obs_info;
}
CV_IMPL void
cvReleaseObsInfo( CvImgObsInfo ** obs_info )
{
IPPI_CALL( icvReleaseObsInfo( obs_info ));
}
CV_IMPL void
cvUniformImgSegm( CvImgObsInfo * obs_info, CvEHMM * hmm )
{
IPPI_CALL( icvUniformImgSegm( obs_info, hmm ));
}
CV_IMPL void
cvInitMixSegm( CvImgObsInfo ** obs_info_array, int num_img, CvEHMM * hmm )
{
IPPI_CALL( icvInitMixSegm( obs_info_array, num_img, hmm ));
}
CV_IMPL void
cvEstimateHMMStateParams( CvImgObsInfo ** obs_info_array, int num_img, CvEHMM * hmm )
{
IPPI_CALL( icvEstimateHMMStateParams( obs_info_array, num_img, hmm ));
}
CV_IMPL void
cvEstimateTransProb( CvImgObsInfo ** obs_info_array, int num_img, CvEHMM * hmm )
{
IPPI_CALL( icvEstimateTransProb( obs_info_array, num_img, hmm ));
}
CV_IMPL void
cvEstimateObsProb( CvImgObsInfo * obs_info, CvEHMM * hmm )
{
IPPI_CALL( icvEstimateObsProb( obs_info, hmm ));
}
CV_IMPL float
cvEViterbi( CvImgObsInfo * obs_info, CvEHMM * hmm )
{
if( (obs_info == NULL) || (hmm == NULL) )
CV_Error( CV_BadDataPtr, "Null pointer." );
return icvEViterbi( obs_info, hmm );
}
CV_IMPL void
cvMixSegmL2( CvImgObsInfo ** obs_info_array, int num_img, CvEHMM * hmm )
{
IPPI_CALL( icvMixSegmL2( obs_info_array, num_img, hmm ));
}
/* End of file */