PCA_Train_Result* PCANet_train(vector<cv::Mat>& InImg, PCANet* PcaNet, bool is_extract_feature){
assert(PcaNet->NumFilters.size() == PcaNet->NumStages);
PCA_Train_Result* train_result = new PCA_Train_Result;
PCA_Out_Result* out_result = new PCA_Out_Result;
PCA_Out_Result* temp;
out_result->OutImg = InImg;
int img_length = InImg.size();
for(int i=0; i<img_length; i++)
out_result->OutImgIdx.push_back(i);
int64 e1 = cv::getTickCount();
int64 eo1, eo2, eb1, eb2;
for(int s=0; s<PcaNet->NumStages; s++){
eb1 = cv::getTickCount();
cout << " Computing PCA filter bank and its outputs at stage " << s << "..." << endl;
train_result->Filters.push_back(PCA_FilterBank(out_result->OutImg, PcaNet->PatchSize, PcaNet->NumFilters[s]));
eb2 = cv::getTickCount();
cout <<" stage"<<s<<" PCA_FilterBank time: "<<(eb2 - eb1)/ cv::getTickFrequency()<<endl;
eo1 = cv::getTickCount();
if(s != PcaNet->NumStages - 1){
temp = PCA_output(out_result->OutImg, out_result->OutImgIdx, PcaNet->PatchSize,
PcaNet->NumFilters[s], train_result->Filters[s], omp_get_num_procs());
delete out_result;
out_result = temp;
}
eo2 = cv::getTickCount();
cout <<" stage"<<s<<" output time: "<<(eo2 - eo1)/ cv::getTickFrequency()<<endl;
}
int64 e2 = cv::getTickCount();
double time = (e2 - e1)/ cv::getTickFrequency();
cout <<"\n totle FilterBank time: "<<time<<endl;
InImg.clear();
vector<cv::Mat>().swap(InImg);
vector<cv::Mat> tempF;
int end = PcaNet->NumStages - 1;
int outIdx_length = out_result->OutImgIdx.size();
if(is_extract_feature){
vector<cv::Mat>::const_iterator first = out_result->OutImg.begin();
vector<cv::Mat>::const_iterator last = out_result->OutImg.begin();
vector<cv::Mat> features;
Hashing_Result* hashing_r;
int coreNum = omp_get_num_procs();//»ñµÃŽŠÀíÆ÷žöÊý
e1 = cv::getTickCount();
# pragma omp parallel for default(none) num_threads(coreNum) private(temp, hashing_r) shared(features, out_result, PcaNet, first, last, outIdx_length, img_length, train_result, end)
for(int i=0; i<img_length; i++){
vector<cv::Mat> subInImg(first + i * PcaNet->NumFilters[end], last + (i + 1) * PcaNet->NumFilters[end]);
vector<int> subIdx;
/*for(int j=0; j<outIdx_length; j++){
if(out_result->OutImgIdx[j] == i) subIdx.push_back(1);
else subIdx.push_back(0);
}*/
for(int j=0; j< PcaNet->NumFilters[end]; j++)
subIdx.push_back(j);
temp = PCA_output(subInImg, subIdx, PcaNet->PatchSize,
PcaNet->NumFilters[end], train_result->Filters[end], 2);
hashing_r = HashingHist(PcaNet, temp->OutImgIdx, temp->OutImg);
#pragma omp critical
{
features.push_back(hashing_r->Features);
train_result->feature_idx.push_back(out_result->OutImgIdx[i]);
}
delete hashing_r;
delete temp;
subIdx.clear();
vector<int>().swap(subIdx);
}
e2 = cv::getTickCount();
time = (e2 - e1)/ cv::getTickFrequency();
cout <<"\n hasing time: "<<time<<endl;
//out_result->OutImg.clear();
//vector<cv::Mat>().swap(out_result->OutImg);
delete out_result;
int size = features.size();
if(size > 0){
train_result->Features.create(0, features[0].cols, features[0].type());
for(int i=0 ;i<size; i++){
train_result->Features.push_back(features[i]);
}
/*
train_result->Features = features[0];
for(int i=1 ;i<size; i++){
vconcat(train_result->Features, features[i], train_result->Features);
}*/
}
features.clear();
vector<cv::Mat>().swap(features);
}
//if(temp != NULL)
// delete temp;
return train_result;
}
void Objectness::PCANet_pred(Mat img, vector<Mat> &testImg, vector<Rect> &bboxes, PCA_Train_Result* result, PCANet pcaNet){
// cout << "\n ====== PCANet Testing ======= \n" << endl;
int testSIze = testImg.size();
// cout << testSIze << endl;
Hashing_Result* hashing_r;
PCA_Out_Result *out;
PCA_Out_Result *res;
int coreNum = omp_get_num_procs();
#pragma omp parallel for default(none) num_threads(coreNum) private(out, res,hashing_r) shared(mySVM,libModel, pcaNet, testSIze, testImg, result, bboxes,img,color)
for (int i = 0; i < testSIze; i++){
out = new PCA_Out_Result;
out->OutImgIdx.push_back(0);
out->OutImg.push_back(testImg[i]);
PCA_output(res,out->OutImg, out->OutImgIdx, pcaNet.PatchSize,
pcaNet.NumFilters[0], result->Filters[0], 2);
for (int j = 1; j < pcaNet.NumFilters[1]; j++)
res->OutImgIdx.push_back(j);
delete out;
out = new PCA_Out_Result;
PCA_output(out, res->OutImg, res->OutImgIdx, pcaNet.PatchSize,
pcaNet.NumFilters[1], result->Filters[1], 2);
HashingHist(hashing_r,&pcaNet, out->OutImgIdx, out->OutImg);
hashing_r->Features.convertTo(hashing_r->Features, CV_32F);
delete res;
delete out;
// OpenCV SVM prediction
float pred = mySVM.predict(hashing_r->Features/*, true*/);
/*========== liblinear prediction =================*/
//struct feature_node *svmVec;
//svmVec = (struct feature_node *) malloc((hashing_r->Features.cols + 1)*sizeof(struct feature_node));
//double *predictions = new double[hashing_r->Features.rows];
//float *dataPtr = hashing_r->Features.ptr<float>(); // Get data from OpenCV Mat
//double prob_est[2]; // Probability estimation
//int r, c;
//for (r = 0; r < hashing_r->Features.rows; r++)
//{
// for (c = 0; c < hashing_r->Features.cols; c++)
// {
// svmVec[c].index = c + 1; // Index starts from 1; Pre-computed kernel starts from 0
// svmVec[c].value = dataPtr[r*hashing_r->Features.cols + c];
// }
// svmVec[c].index = -1; // End of line
//}
//double pred = predict(libModel, svmVec);
//free(svmVec);
//printf("pred: %f\n", pred);
delete hashing_r;
#pragma omp critical
if (pred < 3){
//cout << imgnames[i] << endl;
Rect box = bboxes[i];
//printf("predict: %f\n", pred);
/*========== save to patch =============*/
//cv::Mat croppedImage;
//img(box).copyTo(croppedImage);
////imshow("asd", croppedImage);
//char* patch_name = new char[30];
//sprintf(patch_name, "%s%d%s", "patch/", k + 1, ".jpg");
//imwrite(patch_name, croppedImage);
//k++;
rectangle(img, cvPoint(box.x, box.y), cvPoint(box.x + box.width, box.y + box.height), color, 1);
// ========= add some text ============
char buf[50] = { 0 };
if (pred == 1)
sprintf_s(buf, 50, "right");
else if (pred == 2)
sprintf_s(buf, 50, "forward");
else
sprintf_s(buf, 50, "left");
string text = buf;
int baseline = 0;
Size textSize = getTextSize(text, fontFace, fontScale, 1, &baseline);
Point2f textOrg2(box.x, box.y + textSize.height + 2);
putText(img, text, textOrg2, fontFace, fontScale, fontColor);
}
}
}
