利用cmake对opencv的资源进行编译，得到opencv_traincascade工程，并设置其为启动项，利用准备好的训练样本，即可训练。
从main（）函数开始
其中的参数都是自己设置的，具体含义
classifier.train( cascadeDirName,//分类器存放地址
vecName,//vec文件名
bgName,//负样本文件名
numPos, numNeg,//numpos表示每一个训练阶段正样本数量，numNeg每一个训练阶段负样本数
precalcValBufSize, precalcIdxBufSize,
numStages,//分类器级数
cascadeParams,
*featureParams[cascadeParams.featureType],
stageParams,
baseFormatSave,
acceptanceRatioBreakValue );
return 0;进入classifier.train()函数中
bool CvCascadeClassifier::train( const string _cascadeDirName,//级联分类器训练
const string _posFilename,
const string _negFilename,
int _numPos, int _numNeg,
int _precalcValBufSize, int _precalcIdxBufSize,
int _numStages,
const CvCascadeParams& _cascadeParams,
const CvFeatureParams& _featureParams,//选择了Harr特征
const CvCascadeBoostParams& _stageParams,
bool baseFormatSave,
double acceptanceRatioBreakValue )
{
// Start recording clock ticks for training time output
const clock_t begin_time = clock();
if( _cascadeDirName.empty() || _posFilename.empty() || _negFilename.empty() )
CV_Error( CV_StsBadArg, "_cascadeDirName or _bgfileName or _vecFileName is NULL" );
string dirName;
if (_cascadeDirName.find_last_of("/\\") == (_cascadeDirName.length() - 1) )
dirName = _cascadeDirName;
else
dirName = _cascadeDirName + '/';
numPos = _numPos;
numNeg = _numNeg;
numStages = _numStages;
if ( !imgReader.create( _posFilename, _negFilename, _cascadeParams.winSize ) )
{
cout << "Image reader can not be created from -vec " << _posFilename
<< " and -bg " << _negFilename << "." << endl;
return false;
}
if ( !load( dirName ) )
{
cascadeParams = _cascadeParams;
featureParams = CvFeatureParams::create(cascadeParams.featureType);
featureParams->init(_featureParams);
stageParams = makePtr<CvCascadeBoostParams>();
*stageParams = _stageParams;
featureEvaluator = CvFeatureEvaluator::create(cascadeParams.featureType);
featureEvaluator->init( featureParams, numPos + numNeg, cascadeParams.winSize );
stageClassifiers.reserve( numStages );
}else{
// Make sure that if model parameters are preloaded, that people are aware of this,
// even when passing other parameters to the training command
cout << "---------------------------------------------------------------------------------" << endl;
cout << "Training parameters are pre-loaded from the parameter file in data folder!" << endl;
cout << "Please empty this folder if you want to use a NEW set of training parameters." << endl;
cout << "---------------------------------------------------------------------------------" << endl;
}//输出参数信息
cout << "PARAMETERS:" << endl;
cout << "cascadeDirName: " << _cascadeDirName << endl;
cout << "vecFileName: " << _posFilename << endl;
cout << "bgFileName: " << _negFilename << endl;
cout << "numPos: " << _numPos << endl;
cout << "numNeg: " << _numNeg << endl;
cout << "numStages: " << numStages << endl;
cout << "precalcValBufSize[Mb] : " << _precalcValBufSize << endl;
cout << "precalcIdxBufSize[Mb] : " << _precalcIdxBufSize << endl;
cout << "acceptanceRatioBreakValue : " << acceptanceRatioBreakValue << endl;
cascadeParams.printAttrs();
stageParams->printAttrs();
featureParams->printAttrs();
int startNumStages = (int)stageClassifiers.size();
if ( startNumStages > 1 )
cout << endl << "Stages 0-" << startNumStages-1 << " are loaded" << endl;
else if ( startNumStages == 1)
cout << endl << "Stage 0 is loaded" << endl;
//requiredLeafFARate:要求的分支虚警率0.0000009536743，接近于0
double requiredLeafFARate = pow( (double) stageParams->maxFalseAlarm, (double) numStages ) /
(double)stageParams->max_depth;
double tempLeafFARate;
for( int i = startNumStages; i < numStages; i++ )//进行每一阶段的分类器训练
{
cout << endl << "===== TRAINING " << i << "-stage =====" << endl;
cout << "<BEGIN" << endl;
if ( !updateTrainingSet( requiredLeafFARate, tempLeafFARate ) )//
//从正负样本集合里挑选出numPos+numNeg个样本到集合CvCascadeImageReader imgReader中,
//个人感觉很多时间耗在这里，因为这里对每一个样本要进行predict，就是说利用当前强分类器已构建的弱分类器进行判断是正样本还是负样本，
//只有被判断是正样本的情况才被加到TrainingSet中（第一次的时候当然是默认都是正样本）。
//所以如果i比较大的时候，在构建负样本的TrainingSet就特别费时。并且把积分图像等计算出放到evaluator的img中。
{
//样本数不够, 退出训练
cout << "Train dataset for temp stage can not be filled. "
"Branch training terminated." << endl;//训练停止条件
break;
}
if( tempLeafFARate <= requiredLeafFARate )
{
cout << "Required leaf false alarm rate achieved. "
"Branch training terminated." << endl;
break;
}
if( (tempLeafFARate <= acceptanceRatioBreakValue) && (acceptanceRatioBreakValue >= 0) ){
cout << "The required acceptanceRatio for the model has been reached to avoid overfitting of trainingdata. "
"Branch training terminated." << endl;
break;
}
Ptr<CvCascadeBoost> tempStage = makePtr<CvCascadeBoost>();
bool isStageTrained = tempStage->train( featureEvaluator,
curNumSamples, _precalcValBufSize, _precalcIdxBufSize,
*stageParams );//训练一个强分类器
cout << "END>" << endl;
if(!isStageTrained)
break;
stageClassifiers.push_back( tempStage );//把训练好的强分类器加入到容器中
// save params
if( i == 0)
{
std::string paramsFilename = dirName + CC_PARAMS_FILENAME;
FileStorage fs( paramsFilename, FileStorage::WRITE);
if ( !fs.isOpened() )
{
cout << "Parameters can not be written, because file " << paramsFilename
<< " can not be opened." << endl;
return false;
}
fs << FileStorage::getDefaultObjectName(paramsFilename) << "{";
writeParams( fs );
fs << "}";
}//以下是保存数据
// save current stage
char buf[10];
sprintf(buf, "%s%d", "stage", i );
string stageFilename = dirName + buf + ".xml";
FileStorage fs( stageFilename, FileStorage::WRITE );
if ( !fs.isOpened() )
{
cout << "Current stage can not be written, because file " << stageFilename
<< " can not be opened." << endl;
return false;
}
fs << FileStorage::getDefaultObjectName(stageFilename) << "{";
tempStage->write( fs, Mat() );
fs << "}";
// Output training time up till now
float seconds = float( clock () - begin_time ) / CLOCKS_PER_SEC;
int days = int(seconds) / 60 / 60 / 24;
int hours = (int(seconds) / 60 / 60) % 24;
int minutes = (int(seconds) / 60) % 60;
int seconds_left = int(seconds) % 60;
cout << "Training until now has taken " << days << " days " << hours << " hours " << minutes << " minutes " << seconds_left <<" seconds." << endl;
}
if(stageClassifiers.size() == 0)
{
cout << "Cascade classifier can't be trained. Check the used training parameters." << endl;
return false;
}
save( dirName + CC_CASCADE_FILENAME, baseFormatSave );
return true;
}进入bool isStageTrained = tempStage->train（）//训练强分类器
bool CvCascadeBoost::train( const CvFeatureEvaluator* _featureEvaluator,
int _numSamples,
int _precalcValBufSize, int _precalcIdxBufSize,
const CvCascadeBoostParams& _params )
{
bool isTrained = false;
CV_Assert( !data );
clear();
data = new CvCascadeBoostTrainData( _featureEvaluator, _numSamples,
_precalcValBufSize, _precalcIdxBufSize, _params );//setdata（）,目前暂且知道这里是调用preCalculate计算所有的特征值
CvMemStorage *storage = cvCreateMemStorage();
//创建将驻留在指定存储中的新空序列，弱分类器。
weak = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvBoostTree*), storage );
storage = 0;
set_params( _params );
if ( (_params.boost_type == LOGIT) || (_params.boost_type == GENTLE) )
data->do_responses_copy();
update_weights( 0 );//权重初始化
cout << "+----+---------+---------+" << endl;
cout << "| N | HR | FA |" << endl;
cout << "+----+---------+---------+" << endl;
do
{
CvCascadeBoostTree* tree = new CvCascadeBoostTree;
if( !tree->train( data, subsample_mask, this ) )//训练弱分类器，一个决策树
{
delete tree;
break;
}
cvSeqPush( weak, &tree );//把弱分类器添加到强分类器里面
update_weights( tree );//更新权重。
trim_weights();
if( cvCountNonZero(subsample_mask) == 0 )
break;
}
while( !isErrDesired() && (weak->total < params.weak_count) );//训练出的弱分类器的个数小于参数值：100并且虚警率不小于0.5则训练继续。
if(weak->total > 0)
{
data->is_classifier = true;
data->free_train_data();
isTrained = true;
}
else
clear();
return isTrained;
}tree->train( data, subsample_mask, this ) )//训练弱分类器，一个决策树
CvBoostTree::train( CvDTreeTrainData* _train_data,
const CvMat* _subsample_idx, CvBoost* _ensemble )
{
clear();
ensemble = _ensemble;
data = _train_data;
data->shared = true;
return do_train( _subsample_idx );//传递的是：_subsample_idx子样本索引
}进入do_train() ，在这里获得弱分类器的是生成一个决策树（李航.统计学习方法中有关决策树的论述）
bool CvDTree::do_train( const CvMat* _subsample_idx )//训练弱分类器
{
//1）对于每个特征 f，计算所有训练样本的特征值，并将其排序。
// 扫描一遍排好序的特征值，对排好序的表中的每个元素，计算下面四个值：
// 全部人脸样本的权重的和t1；
// 全部非人脸样本的权重的和t0；
// 在此元素之前的人脸样本的权重的和s1；
// 在此元素之前的非人脸样本的权重的和s0；
// 2）最终求得每个元素的分类误差
// 在表中寻找r值最小的元素，则该元素作为最优阈值。有了该阈值，我们的第一个最优弱分类器就诞生了。<span style="white-space:pre"> </span>//以上的弱分类器的寻找是在网上比较常见的描述但是在opencv_traincascade中，使用的决策树（）
bool result = false;
CV_FUNCNAME( "CvDTree::do_train" );
__BEGIN__;
root = data->subsample_data( _subsample_idx );//子样本数据
CV_CALL( try_split_node(root));//尝试分裂节点，生成树
if( root->split )
{
CV_Assert( root->left );
CV_Assert( root->right );
if( data->params.cv_folds > 0 )
CV_CALL( prune_cv() );//修建生成树
if( !data->shared )
data->free_train_data();
result = true;
}
__END__;
return result;
}CV_CALL( try_split_node(root));//尝试分裂节点，生成树
void CvDTree::try_split_node( CvDTreeNode* node )//分裂节点，生成树
{
CvDTreeSplit* best_split = 0;
int i, n = node->sample_count, vi;
bool can_split = true;
double quality_scale;
calc_node_value( node );//计算节点值（就是一个决策树），根节点值，最优特征
if( node->sample_count <= data->params.min_sample_count ||
node->depth >= data->params.max_depth )
can_split = false;//根据条件停止生成树
if( can_split && data->is_classifier )
{
// check if we have a "pure" node,
// we assume that cls_count is filled by calc_node_value()
int* cls_count = data->counts->data.i;
int nz = 0, m = data->get_num_classes();
for( i = 0; i < m; i++ )
nz += cls_count[i] != 0;
if( nz == 1 ) // there is only one class
can_split = false;
}
else if( can_split )
{
if( sqrt(node->node_risk)/n < data->params.regression_accuracy )
can_split = false;
}
if( can_split )
{
best_split = find_best_split(node);
// TODO: check the split quality ...
node->split = best_split;
}
if( !can_split || !best_split )
{
data->free_node_data(node);
return;
}
quality_scale = calc_node_dir( node );
if( data->params.use_surrogates )
{
// find all the surrogate splits
// and sort them by their similarity to the primary one
//找到所有的代理拆分，并通过它们与主要的相似性对它们进行排序
for( vi = 0; vi < data->var_count; vi++ )
{
CvDTreeSplit* split;
int ci = data->get_var_type(vi);
if( vi == best_split->var_idx )
continue;
if( ci >= 0 )
split = find_surrogate_split_cat( node, vi );
else
split = find_surrogate_split_ord( node, vi );
if( split )
{
// insert the split
//插入分裂
CvDTreeSplit* prev_split = node->split;
split->quality = (float)(split->quality*quality_scale);
while( prev_split->next &&
prev_split->next->quality > split->quality )
prev_split = prev_split->next;
split->next = prev_split->next;
prev_split->next = split;
}
}
}
split_node_data( node );
try_split_node( node->left );//左节点
try_split_node( node->right );//右节点
}calc_node_value( node );//计算节点值（就是一个决策树），根节点值，最优特征
void CvDTree::calc_node_value( CvDTreeNode* node )//得到一个决策树，一个弱分类器，根节点就是最优特征
{
int i, j, k, n = node->sample_count, cv_n = data->params.cv_folds;
int m = data->get_num_classes();
int base_size = data->is_classifier ? m*cv_n*sizeof(int) : 2*cv_n*sizeof(double)+cv_n*sizeof(int);
int ext_size = n*(sizeof(int) + (data->is_classifier ? sizeof(int) : sizeof(int)+sizeof(float)));
cv::AutoBuffer<uchar> inn_buf(base_size + ext_size);
uchar* base_buf = (uchar*)inn_buf;
uchar* ext_buf = base_buf + base_size;
int* cv_labels_buf = (int*)ext_buf;
const int* cv_labels = data->get_cv_labels(node, cv_labels_buf);
if( data->is_classifier )
{
// in case of classification tree:决策树中的分类树
// * node value is the label of the class that has the largest weight in the node.
// * node risk is the weighted number of misclassified samples,
// * j-th cross-validation fold value and risk are calculated as above,
// but using the samples with cv_labels(*)!=j.
// * j-th cross-validation fold error is calculated as the weighted number of
// misclassified samples with cv_labels(*)==j.
// compute the number of instances of each class
int* cls_count = data->counts->data.i;
int* responses_buf = cv_labels_buf + n;
const int* responses = data->get_class_labels(node, responses_buf);
int* cv_cls_count = (int*)base_buf;
double max_val = -1, total_weight = 0;
int max_k = -1;
double* priors = data->priors_mult->data.db;
for( k = 0; k < m; k++ )
cls_count[k] = 0;
if( cv_n == 0 )
{
for( i = 0; i < n; i++ )
cls_count[responses[i]]++;
}
else
{
for( j = 0; j < cv_n; j++ )
for( k = 0; k < m; k++ )
cv_cls_count[j*m + k] = 0;
for( i = 0; i < n; i++ )
{
j = cv_labels[i]; k = responses[i];
cv_cls_count[j*m + k]++;
}
for( j = 0; j < cv_n; j++ )
for( k = 0; k < m; k++ )
cls_count[k] += cv_cls_count[j*m + k];
}
if( data->have_priors && node->parent == 0 )
{
// compute priors_mult from priors, take the sample ratio into account.
double sum = 0;
for( k = 0; k < m; k++ )
{
int n_k = cls_count[k];
priors[k] = data->priors->data.db[k]*(n_k ? 1./n_k : 0.);
sum += priors[k];
}
sum = 1./sum;
for( k = 0; k < m; k++ )
priors[k] *= sum;
}
for( k = 0; k < m; k++ )
{
double val = cls_count[k]*priors[k];
total_weight += val;
if( max_val < val )
{
max_val = val;
max_k = k;
}
}
node->class_idx = max_k;
node->value = data->cat_map->data.i[
data->cat_ofs->data.i[data->cat_var_count] + max_k];
node->node_risk = total_weight - max_val;
for( j = 0; j < cv_n; j++ )
{
double sum_k = 0, sum = 0, max_val_k = 0;
max_val = -1; max_k = -1;
for( k = 0; k < m; k++ )
{
double w = priors[k];
double val_k = cv_cls_count[j*m + k]*w;
double val = cls_count[k]*w - val_k;
sum_k += val_k;
sum += val;
if( max_val < val )
{
max_val = val;
max_val_k = val_k;
max_k = k;
}
}
node->cv_Tn[j] = INT_MAX;
node->cv_node_risk[j] = sum - max_val;
node->cv_node_error[j] = sum_k - max_val_k;
}
}
else
{
// in case of regression tree:决策树种的回归树
// * node value is 1/n*sum_i(Y_i), where Y_i is i-th response,
// n is the number of samples in the node.节点样本数
// * node risk is the sum of squared errors: sum_i((Y_i - <node_value>)^2);平方误差
// * j-th cross-validation fold value and risk are calculated as above,较差验证折叠值和平方误差计算
// but using the samples with cv_labels(*)!=j.
// * j-th cross-validation fold error is calculated
// using samples with cv_labels(*)==j as the test subset:
// error_j = sum_(i,cv_labels(i)==j)((Y_i - <node_value_j>)^2),
// where node_value_j is the node value calculated
// as described in the previous bullet, and summation is done
// over the samples with cv_labels(*)==j.
double sum = 0, sum2 = 0;
float* values_buf = (float*)(cv_labels_buf + n);
int* sample_indices_buf = (int*)(values_buf + n);
const float* values = data->get_ord_responses(node, values_buf, sample_indices_buf);
double *cv_sum = 0, *cv_sum2 = 0;
int* cv_count = 0;
if( cv_n == 0 )
{
for( i = 0; i < n; i++ )
{
double t = values[i];
sum += t;
sum2 += t*t;
}
}
else
{
cv_sum = (double*)base_buf;
cv_sum2 = cv_sum + cv_n;
cv_count = (int*)(cv_sum2 + cv_n);
for( j = 0; j < cv_n; j++ )
{
cv_sum[j] = cv_sum2[j] = 0.;
cv_count[j] = 0;
}
for( i = 0; i < n; i++ )
{
j = cv_labels[i];
double t = values[i];
double s = cv_sum[j] + t;
double s2 = cv_sum2[j] + t*t;
int nc = cv_count[j] + 1;
cv_sum[j] = s;
cv_sum2[j] = s2;
cv_count[j] = nc;
}
for( j = 0; j < cv_n; j++ )
{
sum += cv_sum[j];
sum2 += cv_sum2[j];
}
}
node->node_risk = sum2 - (sum/n)*sum;//平方误差总和
node->value = sum/n;
for( j = 0; j < cv_n; j++ )
{
double s = cv_sum[j], si = sum - s;
double s2 = cv_sum2[j], s2i = sum2 - s2;
int c = cv_count[j], ci = n - c;
double r = si/MAX(ci,1);
node->cv_node_risk[j] = s2i - r*r*ci;
node->cv_node_error[j] = s2 - 2*r*s + c*r*r;
node->cv_Tn[j] = INT_MAX;
}
}
} do
{
CvCascadeBoostTree* tree = new CvCascadeBoostTree;
if( !tree->train( data, subsample_mask, this ) )//训练弱分类器，一个决策树
{
delete tree;
break;
}
cvSeqPush( weak, &tree );//把弱分类器添加到强分类器里面
update_weights( tree );//更新权重。
trim_weights();
if( cvCountNonZero(subsample_mask) == 0 )
break;
}
while( !isErrDesired() && (weak->total < params.weak_count) )进入权重更新：update_weights(tree),在更新权重的过程中，里面有四个Adaboost的更新算法discrete Adaboost、Real Adaboost、Gentle Adaboost和 LogitBoost。其中的算法描述在《基于子空间的人脸识别》中有详细介绍，网络上的描述内容不详细 。
void CvCascadeBoost::update_weights( CvBoostTree* tree )
{
int n = data->sample_count;//1000个训练样本
double sumW = 0.;
int step = 0;
float* fdata = 0;
int *sampleIdxBuf;
const int* sampleIdx = 0;
int inn_buf_size = ((params.boost_type == LOGIT) || (params.boost_type == GENTLE) ? n*sizeof(int) : 0) +
( !tree ? n*sizeof(int) : 0 );
cv::AutoBuffer<uchar> inn_buf(inn_buf_size);
uchar* cur_inn_buf_pos = (uchar*)inn_buf;
if ( (params.boost_type == LOGIT) || (params.boost_type == GENTLE) )
{
step = CV_IS_MAT_CONT(data->responses_copy->type) ?
1 : data->responses_copy->step / CV_ELEM_SIZE(data->responses_copy->type);
fdata = data->responses_copy->data.fl;
sampleIdxBuf = (int*)cur_inn_buf_pos; cur_inn_buf_pos = (uchar*)(sampleIdxBuf + n);
sampleIdx = data->get_sample_indices( data->data_root, sampleIdxBuf );
}
CvMat* buf = data->buf;
size_t length_buf_row = data->get_length_subbuf();
if( !tree ) // before training the first tree, initialize weights and other parameters初始化权重和其他参数
{
int* classLabelsBuf = (int*)cur_inn_buf_pos; cur_inn_buf_pos = (uchar*)(classLabelsBuf + n);
const int* classLabels = data->get_class_labels(data->data_root, classLabelsBuf);
// in case of logitboost and gentle adaboost each weak tree is a regression tree,
// so we need to convert class labels to floating-point values
double w0 = 1./n;
double p[2] = { 1, 1 };
cvReleaseMat( &orig_response );
cvReleaseMat( &sum_response );
cvReleaseMat( &weak_eval );
cvReleaseMat( &subsample_mask );
cvReleaseMat( &weights );
orig_response = cvCreateMat( 1, n, CV_32S );
weak_eval = cvCreateMat( 1, n, CV_64F );//1*1000的矩阵
subsample_mask = cvCreateMat( 1, n, CV_8U );
weights = cvCreateMat( 1, n, CV_64F );//1*1000的矩阵
subtree_weights = cvCreateMat( 1, n + 2, CV_64F );
if (data->is_buf_16u)
{
unsigned short* labels = (unsigned short*)(buf->data.s + data->data_root->buf_idx*length_buf_row +
data->data_root->offset + (size_t)(data->work_var_count-1)*data->sample_count);
for( int i = 0; i < n; i++ )
{
// save original categorical responses {0,1}, convert them to {-1,1}
orig_response->data.i[i] = classLabels[i]*2 - 1;//
// make all the samples active at start.
// later, in trim_weights() deactivate/reactive again some, if need
subsample_mask->data.ptr[i] = (uchar)1;
// make all the initial weights the same.初始权值相同
weights->data.db[i] = w0*p[classLabels[i]];//weight->data.db = 0.001.
// set the labels to find (from within weak tree learning proc)
// the particular sample weight, and where to store the response.
labels[i] = (unsigned short)i;
}
}
else
{
int* labels = buf->data.i + data->data_root->buf_idx*length_buf_row +
data->data_root->offset + (size_t)(data->work_var_count-1)*data->sample_count;
for( int i = 0; i < n; i++ )
{
// save original categorical responses {0,1}, convert them to {-1,1}
orig_response->data.i[i] = classLabels[i]*2 - 1;
subsample_mask->data.ptr[i] = (uchar)1;
weights->data.db[i] = w0*p[classLabels[i]];
labels[i] = i;
}
}
if( params.boost_type == LOGIT )
{
sum_response = cvCreateMat( 1, n, CV_64F );
for( int i = 0; i < n; i++ )
{
sum_response->data.db[i] = 0;
fdata[sampleIdx[i]*step] = orig_response->data.i[i] > 0 ? 2.f : -2.f;
}
// in case of logitboost each weak tree is a regression tree.
// the target function values are recalculated for each of the trees
data->is_classifier = false;
}
else if( params.boost_type == GENTLE )
{
for( int i = 0; i < n; i++ )
fdata[sampleIdx[i]*step] = (float)orig_response->data.i[i];
data->is_classifier = false;
}
}
else
{
// at this moment, for all the samples that participated in the training of the most
// recent weak classifier we know the responses. For other samples we need to compute them
if( have_subsample )
{
// invert the subsample mask
cvXorS( subsample_mask, cvScalar(1.), subsample_mask );
// run tree through all the non-processed samples
for( int i = 0; i < n; i++ )
if( subsample_mask->data.ptr[i] )
{
weak_eval->data.db[i] = ((CvCascadeBoostTree*)tree)->predict( i )->value;//有可能是遍历所有的训练样本
//得到弱分类器的分类结果
}
}
// now update weights and other parameters for each type of boosting
if( params.boost_type == DISCRETE )
{
// Discrete AdaBoost:
// weak_eval[i] (=f(x_i)) is in {-1,1}弱分类器的形式和取值
// err = sum(w_i*(f(x_i) != y_i))/sum(w_i);y_i对应于（x_i, y_i）表示训练样本x_i表示Haar特征表示，y_i表示每个样本对应的标签
// C = log((1-err)/err)加权系数
// w_i *= exp(C*(f(x_i) != y_i))更新权值
double C, err = 0.;
double scale[] = { 1., 0. };
for( int i = 0; i < n; i++ )
{
double w = weights->data.db[i];//在update_weights（0）中.db[i]的值是0.001
sumW += w;//暂时理解为权值和
err += w*(weak_eval->data.db[i] != orig_response->data.i[i]);
//weak_eval->data.db[i]表示分类器的输出结果f（x_i）,是构造决策树之后遍历所有样本。
//orig_response->data.i[i]表示训练样本的标签{1, -1}正负样本
//如果相等，表示检测结果与实际相同（正确检测），如果不等表示误检，w*()表示错误率
}
if( sumW != 0 )
err /= sumW;
C = err = -logRatio( err );
scale[1] = exp(err);
sumW = 0;
for( int i = 0; i < n; i++ )
{
double w = weights->data.db[i]*
scale[weak_eval->data.db[i] != orig_response->data.i[i]];//增大分类错误样本权重，不等scale[0] = 1,
//较小分类正确样本权重相等scale[1] = 0；
sumW += w;
weights->data.db[i] = w;//权值更新
}
tree->scale( C );//node 放缩
}
else if( params.boost_type == REAL )
{
// Real AdaBoost:是实数输出的Adaboost算法
// weak_eval[i] = f(x_i) = 0.5*log(p(x_i)/(1-p(x_i))), p(x_i)=P(y=1|x_i)；弱分类器表示形式
// w_i *= exp(-y_i*f(x_i))更新权重
for( int i = 0; i < n; i++ )
weak_eval->data.db[i] *= -orig_response->data.i[i];
cvExp( weak_eval, weak_eval );
for( int i = 0; i < n; i++ )
{
double w = weights->data.db[i]*weak_eval->data.db[i];
sumW += w;
weights->data.db[i] = w;
}
}
else if( params.boost_type == LOGIT )
{
// LogitBoost:
// weak_eval[i] = f(x_i) in [-z_max,z_max]
// sum_response = F(x_i).
// F(x_i) += 0.5*f(x_i)
// p(x_i) = exp(F(x_i))/(exp(F(x_i)) + exp(-F(x_i))=1/(1+exp(-2*F(x_i)))
// reuse weak_eval: weak_eval[i] <- p(x_i)
// w_i = p(x_i)*1(1 - p(x_i))
// z_i = ((y_i+1)/2 - p(x_i))/(p(x_i)*(1 - p(x_i)))
// store z_i to the data->data_root as the new target responses
const double lbWeightThresh = FLT_EPSILON;
const double lbZMax = 10.;
for( int i = 0; i < n; i++ )
{
double s = sum_response->data.db[i] + 0.5*weak_eval->data.db[i];
sum_response->data.db[i] = s;
weak_eval->data.db[i] = -2*s;
}
cvExp( weak_eval, weak_eval );
for( int i = 0; i < n; i++ )
{
double p = 1./(1. + weak_eval->data.db[i]);
double w = p*(1 - p), z;
w = MAX( w, lbWeightThresh );
weights->data.db[i] = w;
sumW += w;
if( orig_response->data.i[i] > 0 )
{
z = 1./p;
fdata[sampleIdx[i]*step] = (float)min(z, lbZMax);
}
else
{
z = 1./(1-p);
fdata[sampleIdx[i]*step] = (float)-min(z, lbZMax);
}
}
}
else
{
// Gentle AdaBoost:是基于加性回归模型的Adaboost算法
// weak_eval[i] = f(x_i) in [-1,1]；弱分类器利用x_i和y_i做加权最小二乘回归，得到弱分类器
// w_i *= exp(-y_i*f(x_i))
assert( params.boost_type == GENTLE );//如果为假，输出错误信息，然后调用abort终止程序运行
for( int i = 0; i < n; i++ )
weak_eval->data.db[i] *= -orig_response->data.i[i];//这个循环是：弱分类分类结果
//其中weak_eval表示弱分类器
cvExp( weak_eval, weak_eval );
for( int i = 0; i < n; i++ )
{
double w = weights->data.db[i] * weak_eval->data.db[i];
weights->data.db[i] = w;
sumW += w;
}
}
}
// renormalize weights归一化权重，使得权重之和 = 1
if( sumW > FLT_EPSILON )
{
sumW = 1./sumW;
for( int i = 0; i < n; ++i )
weights->data.db[i] *= sumW;
}
double original_ = *(weights->data.db);//weights.db = 0.001,1/1000样本初始权重
}