void compute_boosting_score_factor(const PartApp &part_app, int pidx, kma::ImageContent *kmaimg,
const vector<PartBBox> &part_samples, const vector<double> &vect_scales,
Factor1d &factor)
{
assert(part_samples.size() == vect_scales.size());
factor.varidx = pidx;
/** add border */
int added_border = 0;
kma::ImageContent *kmaimg_border = add_image_border2(kmaimg, 50, added_border);
cout << "added_border: " << added_border << endl;
AdaBoostClassifier abc;
part_app.loadClassifier(abc, pidx);
int num_samples = part_samples.size();
factor.fval.resize(boost::extents[num_samples]);
for (int sampidx = 0; sampidx < num_samples; ++sampidx) {
PartBBox bbox = part_samples[sampidx];
/** add offset to position of bounding box */
bbox.part_pos(0) += added_border;
bbox.part_pos(1) += added_border;
/** comput features */
vector<float> all_features;
PartBBox adjusted_rect;
// bool bres = part_detect::compute_part_bbox_features_scale(part_app.m_abc_param, part_app.m_window_param,
// bbox, kmaimg_border, pidx, 1,
// all_features, adjusted_rect);
// assert(vect_scales[sampidx] == 1.0);
bool bres = part_detect::compute_part_bbox_features_scale(part_app.m_abc_param, part_app.m_window_param,
bbox, kmaimg_border, pidx, vect_scales[sampidx],
all_features, adjusted_rect);
factor.fval[sampidx] = part_app.m_abc_param.min_score();
/** evaluate classifier */
if (bres) {
float score = abc.evaluateFeaturePoint(all_features, true);
if (score > part_app.m_abc_param.min_score())
factor.fval[sampidx] = score;
}
else {
cout << "warning: failed to compute features" << endl;
}
}// samples
delete kmaimg_border;
}
void ValidateTheThreshold(AdaBoostClassifier& ada,ofstream& f)
{
int i,j;
int falsepos,falseneg;
for(i=0;i<validfacecount;i++) features[i] = 0.0;
for(i=0;i<validfacecount;i++)
for(j=0;j<ada.count;j++)
if(ada.scs[j].Apply(validset[i])!=0)
features[i] += ada.alphas[j];
nth_element(features,features+int(validfacecount*(1-node_det_goal)),features+validfacecount);
ada.thresh = features[int(validfacecount*(1-node_det_goal))];
falseneg = facecount;
for(i=0;i<facecount;i++) falseneg -= ada.ApplyImagePatch(trainset[i]);
falsepos = 0;
for(i=facecount;i<totalcount;i++) falsepos += ada.ApplyImagePatch(trainset[i]);
f<<"-----Use the validation set to determine a threshold ------"<<endl;
f<<ada.thresh<<"\t"<<falseneg<<"\t"<<falsepos<<endl;
}
void ForwardFeatureSelection(AdaBoostClassifier& ada,const int round,const bool update,const int method)
{
int* curresult;
int* indexes;
ofstream f;
ada.Clear();
curresult = new int[totalcount]; ASSERT(curresult!=NULL);
indexes = new int[totalfeatures]; ASSERT(indexes != NULL);
f.open(FFS_log_filename);
//cerr << "training weak classifier " << endl;
TrainWeakClassifiers();
LoadWeakClassifiers(indexes);
FFS_Histogram(ada,round,curresult,f,method);
if(update) AppendAdaBoostClassifier(ada);
delete[] curresult; curresult=NULL;
delete[] indexes; indexes=NULL;
f.flush();
f.close();
}
void FFS_Histogram(AdaBoostClassifier& ada,const int round,int* const curresult,ofstream& f,const int method)
{
int i,j,k;
int* hist_pos; int* hist_neg;
int* tempresult;
int error,minerror,minindex;
int thresh,minthresh;
bool* used;
tempresult = new int[totalcount]; ASSERT(tempresult!=NULL);
hist_pos = new int[round+1]; ASSERT(hist_pos!=NULL);
hist_neg = new int[round+1]; ASSERT(hist_neg!=NULL);
used = new bool[totalfeatures]; ASSERT(used!=NULL);
for(i=0;i<totalfeatures;i++) used[i]=false;
f<<"Begin Histogram FFS"<<endl;
ada.InitToGivenSize(round);
for(i=0;i<totalcount;i++) curresult[i] = 0;
for(i=0;i<round;i++)
{
minerror = INT_MAX; minindex = -1; minthresh = INT_MAX;
for(j=0;j<totalfeatures;j++)
{
if(used[j]) continue;
int temp;
for(k=0;k<totalcount;k++) tempresult[k] = curresult[k] + table[j][k];
BuildHistogram(tempresult,i+2,hist_pos,hist_neg,facecount);
error = temp = totalcount-facecount;
for(k=0;k<=i+1;k++)
{
temp = temp - hist_neg[k] + hist_pos[k];
if(temp<error) { error = temp; thresh = k+1; }
}
if(error<minerror)
{
minerror = error;
minindex = j;
minthresh = thresh;
}
}
used[minindex]=true;
ada.alphas[ada.count] = 1;
ada.scs[ada.count] = classifiers[minindex];
ada.count++;
ada.thresh = REAL(minthresh);
for(j=0;j<totalcount;j++) curresult[j] += table[minindex][j];
f<<i+1<<"\t"<<minindex<<"\t"<<minerror<<"\t"<<classifiers[minindex].thresh<<endl;
}
BuildHistogram(curresult,round+1,hist_pos,hist_neg,facecount);
for(i=1;i<=round;i++) { hist_neg[i] += hist_neg[i-1]; hist_pos[i] += hist_pos[i-1]; }
i=0; while(hist_neg[i]<(totalcount-facecount)*(1-node_fp_goal)) i++;
switch(method)
{
case GOAL_METHOD_MIN_ERROR:
break;
case GOAL_METHOD_VALID_DR:
ValidateTheThreshold(ada,f);
break;
case GOAL_METHOD_FIX_FP:
ada.thresh = i + REAL(0.5);
break;
}
f<<ada.thresh<<"\t"<<hist_pos[i-1]<<"\t"<<totalcount-facecount-hist_neg[i-1]<<endl;
int fn,fp;
f<<"Training Error:"<<endl;
fn = 0; for(i=0;i<facecount;i++) if(ada.ApplyImagePatch(trainset[i])==0) fn++;
fp = 0; for(i=facecount;i<totalcount;i++) if(ada.ApplyImagePatch(trainset[i])==1) fp++;
f<<fn<<"\t"<<fp<<endl;
f<<"Validation Error:"<<endl;
fn = 0; for(i=0;i<validfacecount;i++) if(ada.ApplyImagePatch(validset[i])==0) fn++;
fp = 0; for(i=validfacecount;i<validcount;i++) if(ada.ApplyImagePatch(validset[i])==1) fp++;
f<<fn<<"\t"<<fp<<endl;
f<<"End of Histogram FFS"<<endl;
delete[] hist_pos; hist_pos=NULL;
delete[] hist_neg; hist_neg=NULL;
delete[] tempresult; tempresult=NULL;
}
