// extract patch from images
// !!!!!!coution!!!!!!!
// this function is use for negatime mode!!!!!
void extractPosPatches(std::vector<CPosDataset> &posSet,
std::vector<CPosPatch> &posPatch,
CConfig conf,
const int treeNum,
const CClassDatabase &classDatabase){
cv::Rect tempRect;
std::vector<CPosPatch> tPosPatch(0);//, posPatch(0);
std::vector<std::vector<CPosPatch> > patchPerClass(classDatabase.vNode.size());
int pixNum;
nCk nck;
int classNum = 0;
posPatch.clear();
tempRect.width = conf.p_width;
tempRect.height = conf.p_height;
std::cout << "image num is " << posSet.size();
std::cout << "extracting patch from image" << std::endl;
for(int l = 0;l < posSet.size(); ++l){
for(int j = 0; j < posSet.at(l).img.at(0)->cols - conf.p_width; j += conf.stride){
for(int k = 0; k < posSet.at(l).img.at(0)->rows - conf.p_height; k += conf.stride){
tempRect.x = j;
tempRect.y = k;
pixNum = 0;
int centerDepthFlag = 1;
// detect negative patch
for(int m = j; m < j + conf.p_width; ++m){
for(int n = k; n < k + conf.p_height; ++n){
pixNum += (int)(posSet.at(l).img.at(1)->at<ushort>(n, m));
}
}
// set patch class
classNum = classDatabase.search(posSet.at(l).getParam()->getClassName());//dataSet.at(l).className.at(0));
if(classNum == -1){
std::cout << "class not found!" << std::endl;
exit(-1);
}
//tPatch.setPatch(temp, image.at(l), dataSet.at(l), classNum);
CPosPatch posTemp(&posSet.at(l),tempRect);
if (pixNum > 0 && posSet.at(l).img.at(1)->at<ushort>(k + (conf.p_height / 2) + 1, j + (conf.p_width / 2) + 1 ) != 0){
tPosPatch.push_back(posTemp);
patchPerClass.at(classNum).push_back(posTemp);
} // if
}//x
}//y
}//allimages
std::vector<int> patchNum(patchPerClass.size(),conf.patchRatio);
for(int i = 0; i < patchPerClass.size(); ++i){
if(i == treeNum % patchPerClass.size())
patchNum.at(i) = conf.patchRatio;
else
patchNum.at(i) = conf.patchRatio * conf.acPatchRatio;
}
// choose patch from each image
for(int i = 0; i < patchPerClass.size(); ++i){
if(patchPerClass.at(i).size() > conf.patchRatio){
std::set<int> chosenPatch = nck.generate(patchPerClass.at(i).size(),patchNum.at(i));// conf.patchRatio);//totalPatchNum * conf.patchRatio);
std::set<int>::iterator ite = chosenPatch.begin();
while(ite != chosenPatch.end()){
//std::cout << "this is for debug ite is " << tPosPatch.at(*ite).center << std::endl;
posPatch.push_back(patchPerClass.at(i).at(*ite));
ite++;
}
}else{
std::cout << "can't extruct enough patch" << std::endl;
}
}
return;
}
void CRForest::growATree(const int treeNum) {
// positive, negative dataset
std::vector<CPosDataset> posSet(0);
std::vector<CNegDataset> negSet(0);
// positive, negative patch
std::vector<CPosPatch> posPatch(0);
std::vector<CNegPatch> negPatch(0);
char buffer[256];
std::cout << "tree number " << treeNum << std::endl;
// initialize random seed
boost::mt19937 gen( treeNum * static_cast<unsigned long>(time(NULL)) );
//boost::timer t;
//loadTrainPosFile(conf, posSet);//, gen);
loadTrainObjFile(conf,posSet);
CClassDatabase tempClassDatabase;
// extract pos features and register classDatabase
for(int i = 0; i < posSet.size(); ++i) {
//std::cout << i << std::endl;
//std::cout << posSet.at(i).rgb << std::endl;
// if(posSet.at(i).loadImage(conf) == -1 && conf.learningMode != 2){
// exit(-1);
// }
//posSet.at(i).extractFeatures(conf);
//std::cout << posSet.size() << std::endl;
tempClassDatabase.add(posSet.at(i).getParam()->getClassName(),cv::Size(),0);
}
// std::vector<CPosDataset> tempPosSet(0);
// int currentClass = treeNum % tempClassDatabase.vNode.size();
//std::cout << "okashiina" << std::endl;
// for(int i = 0; i < posSet.size(); ++i){
// if(tempClassDatabase.search(posSet.at(i).getClassName()) == currentClass){
// tempPosSet.push_back(posSet.at(i));
// //std::cout << "teketeke" << std::endl;
// }else{
// negSet.push_back(convertPosToNeg2(posSet.at(i)));
// //std::cout << "negneg" << std::endl;
// }
// }
//posSet = tempPosSet;
loadTrainNegFile(conf, negSet);
std::cout << "dataset loaded" << std::endl;
// initialize class database
//classDatabase.clear();
std::cout << "generating appearance from 3D model!" << std::endl;
// extract pos features and register classDatabase
for(int i = 0; i < posSet.size(); ++i) {
//std::cout << i << std::endl;
//std::cout << posSet.at(i).rgb << std::endl;
//#pragma omp critical
posSet.at(i).loadImage(conf, posSet.at(i).getModelPath(), posSet.at(i).getParam());
// if(imgload == -1 && conf.learningMode != 2){
// std::cout << "can't load image files" << std::endl;
// exit(-1);
// }
posSet.at(i).extractFeatures(conf);
//std::cout << posSet.at(i).img.at(1)->type() << std::endl;
//std::cout << "detayo" << std::endl;
//std::cout << posSet.size() << std::endl;
classDatabase.add(posSet.at(i).getParam()->getClassName(),posSet.at(i).img.at(0)->size(),0);
pBar(i,posSet.size(),50);
}
std::cout << std::endl;
classDatabase.show();
// extract neg features
for(int i = 0; i < negSet.size(); ++i) {
// if(negSet.at(i).getModel() != NULL)
// negSet.at(i).loadImage(conf, negSet.at(i).getModelPath(), posSet.at(i))
negSet.at(i).loadImage(conf);
negSet.at(i).extractFeatures(conf);
}
CRTree *tree = new CRTree(conf.min_sample, conf.max_depth, classDatabase.vNode.size(),this->classDatabase);
std::cout << "tree created" << std::endl;
extractPosPatches(posSet,posPatch,conf,treeNum,this->classDatabase);
extractNegPatches(negSet,negPatch,conf);
std::cout << "extracted pathes" << std::endl;
std::vector<int> patchClassNum(classDatabase.vNode.size(), 0);
for(int j = 0; j < posPatch.size(); ++j)
patchClassNum.at(classDatabase.search(posPatch.at(j).getClassName()))++;
// grow tree
//vTrees.at(treeNum)->growTree(vPatches, 0,0, (float)(vPatches.at(0).size()) / ((float)(vPatches.at(0).size()) + (float)(vPatches.at(1).size())), conf, gen, patchClassNum);
//#pragma omp parallel
tree->growTree(posPatch,negPatch, 0,0, ((float)posPatch.size() / (float)(posPatch.size() + negPatch.size())), conf, patchClassNum);
// cv::namedWindow("test");
// cv::imshow("test", *posSet.at(0).feature.at(3));
// cv::waitKey(0);
// cv::destroyAllWindows();
// save tree
sprintf(buffer, "%s%03d.txt",
conf.treepath.c_str(), treeNum + conf.off_tree);
std::cout << "tree file name is " << buffer << std::endl;
tree->saveTree(buffer);
// save class database
sprintf(buffer, "%s%s%03d.txt",
conf.treepath.c_str(),
conf.classDatabaseName.c_str(), treeNum + conf.off_tree);
std::cout << "write tree data" << std::endl;
classDatabase.write(buffer);
//double time = t.elapsed();
//std::cout << "tree " << treeNum << " calicuration time is " << time << std::endl;
sprintf(buffer, "%s%03d_timeResult.txt",conf.treepath.c_str(), treeNum + conf.off_tree);
std::fstream lerningResult(buffer, std::ios::out);
if(lerningResult.fail()) {
std::cout << "can't write result" << std::endl;
}
lerningResult << time << std::endl;
lerningResult.close();
delete tree;
posPatch.clear();
negPatch.clear();
// for(int i = 0; i< posSet.size(); ++i){
// cv::namedWindow("test");
// cv::imshow("test", *posSet.at(i).img.at(0));
// cv::waitKey(0);
// cv::destroyWindow("test");
// }
// for(int i = 0; i < posSet.size(); ++i){
// posSet.at(i).releaseImage();
// }
// for(int i = 0; i< posSet.size(); ++i){
// cv::namedWindow("test");
// cv::imshow("test", *posSet.at(i).feature.at(0));
// cv::waitKey(0);
// cv::destroyWindow("test");
// }
// for(int i = 0; i < posSet.size(); ++i){
// posSet.at(i).releaseFeatures();
// }
posSet.clear();
negSet.clear();
}
void loadTrainPosFile(CConfig conf, std::vector<CPosDataset> &posSet)
{
std::string traindatafilepath = conf.trainpath + PATH_SEP + conf.traindatafile;
int n_folders;
int n_files;
std::vector<std::string> trainimagefolder;
std::vector<CPosDataset> tempDataSet;
std::string trainDataListPath;
int dataSetNum;
CClassDatabase database;
cv::Point tempPoint;
nCk nck;
posSet.clear();
//read train data folder list
std::ifstream in(traindatafilepath.c_str());
if(!in.is_open()){
std::cout << "train data floder list is not found!" << std::endl;
exit(1);
}
// read train pos folder
in >> n_folders;
std::cout << "number of training data folders : " << n_folders << std::endl;
trainimagefolder.resize(n_folders);
for(int i = 0;i < n_folders; ++i)
in >> trainimagefolder.at(i);
// close train pos data folder list
in.close();
std::cout << "train folders lists : " << std::endl;
//read train file name and grand truth from file
tempDataSet.clear();
for(int i = 0;i < n_folders; ++i){
trainDataListPath
= conf.trainpath + PATH_SEP + trainimagefolder.at(i) + PATH_SEP + conf.traindatalist;
std::cout << trainDataListPath << std::endl;
std::string imageFilePath
= conf.trainpath + PATH_SEP + trainimagefolder.at(i) + PATH_SEP;
std::ifstream trainDataList(trainDataListPath.c_str());
if(!trainDataList.is_open()){
std::cout << "can't read " << trainDataListPath << std::endl;
exit(1);
}
trainDataList >> n_files;
for(int j = 0;j < n_files; ++j){
CPosDataset posTemp;
std::string nameTemp;
// posTemp.angles.clear();
// posTemp.centerPoint.clear();
//read file names
trainDataList >> nameTemp;
posTemp.setRgbImagePath(imageFilePath + nameTemp);
trainDataList >> nameTemp;
posTemp.setDepthImagePath(imageFilePath + nameTemp);
trainDataList >> nameTemp;// dummy
//read class name
std::string tempClassName;
trainDataList >> tempClassName;
//temp.className.push_back(tempClassName);
posTemp.setClassName(tempClassName);
// read image size
cv::Mat tempImage = cv::imread(posTemp.getRgbImagePath(),3);
cv::Size tempSize = cv::Size(tempImage.cols, tempImage.rows);
// set center point
tempPoint = cv::Point(tempImage.cols / 2, tempImage.rows / 2);
posTemp.setCenterPoint(tempPoint);
// registre class param to class database
database.add(posTemp.getParam()->getClassName(), tempSize, 0);
//read angle grand truth
double tempAngle;
trainDataList >> tempAngle;
posTemp.setAngle(tempAngle);
tempDataSet.push_back(posTemp);
}
trainDataList.close();
}
dataSetNum = tempDataSet.size();
int dataOffset = 0;
database.show();
for(int j = 0;j < database.vNode.size(); j++){
std::set<int> chosenData = nck.generate(database.vNode.at(j).instances, conf.imagePerTree);
std::set<int>::iterator ite = chosenData.begin();
while(ite != chosenData.end()){
posSet.push_back(tempDataSet.at(*ite + dataOffset));
ite++;
}
dataOffset += database.vNode.at(j).instances;
}
// std::cout << "show chosen dataset" << std::endl;
// for(int i = 0; i < dataSet.size(); ++i){
// dataSet.at(i).showDataset();
// }
}
void loadTrainFile(CConfig conf, std::vector<CDataset> &dataSet)
{
dataSet.resize(0);
std::string traindatafilepath = conf.trainpath + PATH_SEP + conf.traindatafile;
int n_folders;
int n_files;
std::vector<std::string> trainimagefolder;
CDataset temp;
std::vector<CDataset> tempDataSet;
std::string trainDataListPath;
int dataSetNum;
CClassDatabase database;
cv::Point tempPoint;
nCk nck;
//read train data folder list
std::ifstream in(traindatafilepath.c_str());
if(!in.is_open()){
std::cout << "train data floder list is not found!" << std::endl;
exit(1);
}
//std::cout << "kokomade " << std::endl;
// read folder number
in >> n_folders;
std::cout << "number of training data folder: "
<< n_folders << std::endl;
trainimagefolder.resize(n_folders);
// read train folder name
for(int i = 0;i < n_folders; ++i)
in >> trainimagefolder.at(i);
std::cout << "train image foloder num is " << trainimagefolder.size() << std::endl;
in.close();
//read train file name and grand truth from file
tempDataSet.clear();
for(int i = 0;i < n_folders; ++i){
trainDataListPath
= conf.trainpath + PATH_SEP + trainimagefolder.at(i)
+ PATH_SEP + conf.traindatalist;
std::cout << trainDataListPath << std::endl;
temp.imageFilePath
= conf.trainpath + PATH_SEP + trainimagefolder.at(i) + PATH_SEP;
std::ifstream trainDataList(trainDataListPath.c_str());
if(!trainDataList.is_open()){
std::cout << "can't read " << trainDataListPath << std::endl;
exit(0);
}
trainDataList >> n_files;
for(int j = 0;j < n_files; ++j){
//read file names
trainDataList >> temp.rgbImageName;
trainDataList >> temp.depthImageName;
trainDataList >> temp.maskImageName;
//read bounding box
//trainDataList >> temp.bBox.x;
//trainDataList >> temp.bBox.y;
//trainDataList >> temp.bBox.width;
//trainDataList >> temp.bBox.height;
temp.centerPoint.clear();
//read center point
//trainDataList >> tempPoint.x;//temp.centerPoint.x;
//trainDataList >> tempPoint.y;
//temp.centerPoint.push_back(tempPoint);
//read class name
trainDataList >> temp.className;
database.add(temp.className);
//read angle grand truth
trainDataList >> temp.angle;
//show read data *for debug
//temp.showDataset();
tempDataSet.push_back(temp);
}
trainDataList.close();
}
dataSetNum = tempDataSet.size();
int dataOffset = 0;
database.show();
for(int j = 0;j < database.vNode.size(); j++){
std::set<int> chosenData = nck.generate(database.vNode.at(j).instances, conf.imagePerTree);
std::set<int>::iterator ite = chosenData.begin();
while(ite != chosenData.end()){
dataSet.push_back(tempDataSet.at(*ite + dataOffset));
ite++;
}
dataOffset += database.vNode.at(j).instances;
}
//in.close();
}
