int main(int argc, char** argv)
{
string filter = "Gaussian";
string descriptor = "HAOG";
string database = "CUFSF";
uint count = 0;
vector<string> extraPhotos, photos, sketches;
loadImages(argv[1], photos);
loadImages(argv[2], sketches);
//loadImages(argv[7], extraPhotos);
uint nPhotos = photos.size(),
nSketches = sketches.size(),
nExtra = extraPhotos.size();
uint nTraining = 2*nPhotos/3;
cout << "Read " << nSketches << " sketches." << endl;
cout << "Read " << nPhotos + nExtra << " photos." << endl;
vector<Mat*> sketchesDescriptors(nSketches), photosDescriptors(nPhotos), extraDescriptors(nExtra);
Mat img, temp;
int size=32, delta=16;
#pragma omp parallel for private(img, temp)
for(uint i=0; i<nSketches; i++){
img = imread(sketches[i],0);
sketchesDescriptors[i] = new Mat();
#pragma omp critical
temp = extractDescriptors(img, size, delta, filter, descriptor);
*(sketchesDescriptors[i]) = temp.clone();
}
#pragma omp parallel for private(img, temp)
for(uint i=0; i<nPhotos; i++){
img = imread(photos[i],0);
photosDescriptors[i] = new Mat();
#pragma omp critical
temp = extractDescriptors(img, size, delta, filter, descriptor);
*(photosDescriptors[i]) = temp.clone();
}
#pragma omp parallel for private(img, temp)
for(uint i=0; i<nExtra; i++){
img = imread(extraPhotos[i],0);
extraDescriptors[i] = new Mat();
#pragma omp critical
temp = extractDescriptors(img, size, delta, filter, descriptor);
*(extraDescriptors[i]) = temp.clone();
}
auto seed = unsigned(count);
srand(seed);
random_shuffle(sketchesDescriptors.begin(), sketchesDescriptors.end());
srand(seed);
random_shuffle(photosDescriptors.begin(), photosDescriptors.end());
//training
vector<Mat*> trainingSketchesDescriptors, trainingPhotosDescriptors;
trainingSketchesDescriptors.insert(trainingSketchesDescriptors.end(), sketchesDescriptors.begin(), sketchesDescriptors.begin()+nTraining);
trainingPhotosDescriptors.insert(trainingPhotosDescriptors.end(), photosDescriptors.begin(), photosDescriptors.begin()+nTraining);
//testing
vector<Mat*> testingSketchesDescriptors, testingPhotosDescriptors;
testingSketchesDescriptors.insert(testingSketchesDescriptors.end(), sketchesDescriptors.begin()+nTraining, sketchesDescriptors.end());
testingPhotosDescriptors.insert(testingPhotosDescriptors.end(), photosDescriptors.begin()+nTraining, photosDescriptors.end());
testingPhotosDescriptors.insert(testingPhotosDescriptors.end(), extraDescriptors.begin(), extraDescriptors.end());
PCA pca;
LDA lda;
vector<int> labels;
uint nTestingSketches = testingSketchesDescriptors.size(),
nTestingPhotos = testingPhotosDescriptors.size();
for(uint i=0; i<nTraining; i++){
labels.push_back(i);
}
labels.insert(labels.end(),labels.begin(),labels.end());
//bags
vector<Mat*> testingSketchesDescriptorsBag(nTestingSketches), testingPhotosDescriptorsBag(nTestingPhotos);
for(int b=0; b<200; b++){
vector<int> bag_indexes = gen_bag(154, 0.1);
uint dim = (bag(*(trainingPhotosDescriptors[0]), bag_indexes, 154)).total();
Mat X(dim, 2*nTraining, CV_32F);
#pragma omp parallel for private(temp)
for(uint i=0; i<nTraining; i++){
temp = *(trainingSketchesDescriptors[i]);
temp = bag(temp, bag_indexes, 154);
temp.copyTo(X.col(i));
}
#pragma omp parallel for private(temp)
for(uint i=0; i<nTraining; i++){
temp = *(trainingPhotosDescriptors[i]);
temp = bag(temp, bag_indexes, 154);
temp.copyTo(X.col(i+nTraining));
}
Mat Xs = X(Range::all(), Range(0,nTraining));
Mat Xp = X(Range::all(), Range(nTraining,2*nTraining));
Mat meanX = Mat::zeros(dim, 1, CV_32F), instance;
Mat meanXs = Mat::zeros(dim, 1, CV_32F);
Mat meanXp = Mat::zeros(dim, 1, CV_32F);
// calculate sums
for (int i = 0; i < X.cols; i++) {
instance = X.col(i);
add(meanX, instance, meanX);
}
for (int i = 0; i < Xs.cols; i++) {
instance = Xs.col(i);
add(meanXs, instance, meanXs);
}
for (int i = 0; i < Xp.cols; i++) {
instance = Xp.col(i);
add(meanXp, instance, meanXp);
}
// calculate total mean
meanX.convertTo(meanX, CV_32F, 1.0/static_cast<double>(X.cols));
meanXs.convertTo(meanXs, CV_32F, 1.0/static_cast<double>(Xs.cols));
meanXp.convertTo(meanXp, CV_32F, 1.0/static_cast<double>(Xp.cols));
// subtract the mean of matrix
for(int i=0; i<X.cols; i++) {
Mat c_i = X.col(i);
subtract(c_i, meanX.reshape(1,dim), c_i);
}
for(int i=0; i<Xs.cols; i++) {
Mat c_i = Xs.col(i);
subtract(c_i, meanXs.reshape(1,dim), c_i);
}
for(int i=0; i<Xp.cols; i++) {
Mat c_i = Xp.col(i);
subtract(c_i, meanXp.reshape(1,dim), c_i);
}
if(meanX.total() >= nTraining)
pca(X, Mat(), CV_PCA_DATA_AS_COL, nTraining-1);
else
pca.computeVar(X, Mat(), CV_PCA_DATA_AS_COL, .99);
Mat W1 = pca.eigenvectors.t();
Mat ldaData = (W1.t()*X).t();
lda.compute(ldaData, labels);
Mat W2 = lda.eigenvectors();
W2.convertTo(W2, CV_32F);
Mat projectionMatrix = (W2.t()*W1.t()).t();
//testing
#pragma omp parallel for private(temp)
for(uint i=0; i<nTestingSketches; i++){
temp = *(testingSketchesDescriptors[i]);
temp = bag(temp, bag_indexes, 154);
temp = projectionMatrix.t()*(temp-meanX);
if(b==0){
testingSketchesDescriptorsBag[i] = new Mat();
*(testingSketchesDescriptorsBag[i]) = temp.clone();
}
else{
vconcat(*(testingSketchesDescriptorsBag[i]), temp, *(testingSketchesDescriptorsBag[i]));
}
}
#pragma omp parallel for private(temp)
for(uint i=0; i<nTestingPhotos; i++){
temp = *(testingPhotosDescriptors[i]);
temp = bag(temp, bag_indexes, 154);
temp = projectionMatrix.t()*(temp-meanX);
if(b==0){
testingPhotosDescriptorsBag[i] = new Mat();
*(testingPhotosDescriptorsBag[i]) = temp.clone();
}
else{
vconcat(*(testingPhotosDescriptorsBag[i]), temp, *(testingPhotosDescriptorsBag[i]));
}
}
}
Mat distancesChi = Mat::zeros(nTestingSketches,nTestingPhotos,CV_64F);
Mat distancesL2 = Mat::zeros(nTestingSketches,nTestingPhotos,CV_64F);
Mat distancesCosine = Mat::zeros(nTestingSketches,nTestingPhotos,CV_64F);
#pragma omp parallel for
for(uint i=0; i<nTestingSketches; i++){
for(uint j=0; j<nTestingPhotos; j++){
distancesChi.at<double>(i,j) = chiSquareDistance(*(testingSketchesDescriptorsBag[i]),*(testingPhotosDescriptorsBag[j]));
distancesL2.at<double>(i,j) = norm(*(testingSketchesDescriptorsBag[i]),*(testingPhotosDescriptorsBag[j]));
distancesCosine.at<double>(i,j) = abs(1-cosineDistance(*(testingSketchesDescriptorsBag[i]),*(testingPhotosDescriptorsBag[j])));
}
}
string file1name = "kernel-drs-" + descriptor + database + to_string(nTraining) + string("chi") + to_string(count) + string(".xml");
string file2name = "kernel-drs-" + descriptor + database + to_string(nTraining) + string("l2") + to_string(count) + string(".xml");
string file3name = "kernel-drs-" + descriptor + database + to_string(nTraining) + string("cosine") + to_string(count) + string(".xml");
FileStorage file1(file1name, FileStorage::WRITE);
FileStorage file2(file2name, FileStorage::WRITE);
FileStorage file3(file3name, FileStorage::WRITE);
file1 << "distanceMatrix" << distancesChi;
file2 << "distanceMatrix" << distancesL2;
file3 << "distanceMatrix" << distancesCosine;
file1.release();
file2.release();
file3.release();
return 0;
}
int main(int argc, char** argv)
{
string filter = "Gaussian";
string descriptor = "SIFT";
string database = "Forensic-extra";
uint count = 1;
vector<string> extraPhotos, photos, sketches;
loadImages(argv[5], photos);
loadImages(argv[6], sketches);
loadImages(argv[7], extraPhotos);
uint nPhotos = photos.size(),
nSketches = sketches.size(),
nExtra = extraPhotos.size();
uint nTraining = 2*nPhotos/3;
cout << "Read " << nSketches << " sketches." << endl;
cout << "Read " << nPhotos + nExtra << " photos." << endl;
vector<Mat*> sketchesDescriptors(nSketches), photosDescriptors(nPhotos), extraDescriptors(nExtra);
Mat img, temp;
int size=32, delta=16;
#pragma omp parallel for private(img, temp)
for(uint i=0; i<nSketches; i++){
img = imread(sketches[i],0);
sketchesDescriptors[i] = new Mat();
#pragma omp critical
temp = extractDescriptors(img, size, delta, filter, descriptor);
*(sketchesDescriptors[i]) = temp.clone();
}
#pragma omp parallel for private(img, temp)
for(uint i=0; i<nPhotos; i++){
img = imread(photos[i],0);
photosDescriptors[i] = new Mat();
#pragma omp critical
temp = extractDescriptors(img, size, delta, filter, descriptor);
*(photosDescriptors[i]) = temp.clone();
}
#pragma omp parallel for private(img, temp)
for(uint i=0; i<nExtra; i++){
img = imread(extraPhotos[i],0);
extraDescriptors[i] = new Mat();
#pragma omp critical
temp = extractDescriptors(img, size, delta, filter, descriptor);
*(extraDescriptors[i]) = temp.clone();
}
auto seed = unsigned(count);
srand(seed);
random_shuffle(sketchesDescriptors.begin(), sketchesDescriptors.end());
srand(seed);
random_shuffle(photosDescriptors.begin(), photosDescriptors.end());
//training
vector<Mat*> trainingSketchesDescriptors1, trainingPhotosDescriptors1,
trainingSketchesDescriptors2, trainingPhotosDescriptors2;
trainingSketchesDescriptors1.insert(trainingSketchesDescriptors1.end(), sketchesDescriptors.begin(), sketchesDescriptors.begin()+nTraining/2);
trainingPhotosDescriptors1.insert(trainingPhotosDescriptors1.end(), photosDescriptors.begin(), photosDescriptors.begin()+nTraining/2);
trainingSketchesDescriptors2.insert(trainingSketchesDescriptors2.end(), sketchesDescriptors.begin()+nTraining/2, sketchesDescriptors.begin()+nTraining);
trainingPhotosDescriptors2.insert(trainingPhotosDescriptors2.end(), photosDescriptors.begin()+nTraining/2, photosDescriptors.begin()+nTraining);
uint nTraining1 = trainingPhotosDescriptors1.size(),
nTraining2 = trainingPhotosDescriptors2.size();
//testing
vector<Mat*> testingSketchesDescriptors, testingPhotosDescriptors;
testingSketchesDescriptors.insert(testingSketchesDescriptors.end(), sketchesDescriptors.begin()+nTraining, sketchesDescriptors.end());
testingPhotosDescriptors.insert(testingPhotosDescriptors.end(), photosDescriptors.begin()+nTraining, photosDescriptors.end());
testingPhotosDescriptors.insert(testingPhotosDescriptors.end(), extraDescriptors.begin(), extraDescriptors.end());
uint nTestingSketches = testingSketchesDescriptors.size(),
nTestingPhotos = testingPhotosDescriptors.size();
PCA pca;
LDA lda;
vector<int> labels;
for(uint i=0; i<nTraining2; i++){
labels.push_back(i);
}
labels.insert(labels.end(),labels.begin(),labels.end());
//bags
vector<Mat*> testingSketchesDescriptorsBag(nTestingSketches), testingPhotosDescriptorsBag(nTestingPhotos),
trainingPhotosDescriptors1Temp(nTraining1), trainingSketchesDescriptors1Temp(nTraining1);
for(int b=0; b<30; b++){
vector<int> bag_indexes = gen_bag(154, 0.1);
#pragma omp parallel for private(temp)
for(uint i=0; i<nTraining1; i++){
temp = *(trainingSketchesDescriptors1[i]);
temp = bag(temp, bag_indexes, 154);
trainingSketchesDescriptors1Temp[i] = new Mat();
*(trainingSketchesDescriptors1Temp[i]) = temp.clone();
}
#pragma omp parallel for private(temp)
for(uint i=0; i<nTraining1; i++){
temp = *(trainingPhotosDescriptors1[i]);
temp = bag(temp, bag_indexes, 154);
trainingPhotosDescriptors1Temp[i] = new Mat();
*(trainingPhotosDescriptors1Temp[i]) = temp.clone();
}
Kernel k(trainingPhotosDescriptors1Temp, trainingSketchesDescriptors1Temp);
k.compute();
uint dim = (k.projectGallery(bag(*(trainingPhotosDescriptors1[0]), bag_indexes, 154))).total();
Mat X(dim, 2*nTraining2, CV_32F);
#pragma omp parallel for private(temp)
for(uint i=0; i<nTraining2; i++){
temp = *(trainingSketchesDescriptors2[i]);
temp = bag(temp, bag_indexes, 154);
temp = k.projectProbe(temp);
temp.copyTo(X.col(i));
}
#pragma omp parallel for private(temp)
for(uint i=0; i<nTraining2; i++){
temp = *(trainingPhotosDescriptors2[i]);
temp = bag(temp, bag_indexes, 154);
temp = k.projectGallery(temp);
temp.copyTo(X.col(i+nTraining2));
}
Mat meanX = Mat::zeros(dim, 1, CV_32F), instance;
// calculate sums
for (int i = 0; i < X.cols; i++) {
instance = X.col(i);
add(meanX, instance, meanX);
}
// calculate total mean
meanX.convertTo(meanX, CV_32F, 1.0/static_cast<double>(X.cols));
// subtract the mean of matrix
for(int i=0; i<X.cols; i++) {
Mat c_i = X.col(i);
subtract(c_i, meanX.reshape(1,dim), c_i);
}
pca.computeVar(X, Mat(), CV_PCA_DATA_AS_COL, .99);
Mat W1 = pca.eigenvectors.t();
Mat ldaData = (W1.t()*X).t();
lda.compute(ldaData, labels);
Mat W2 = lda.eigenvectors();
W2.convertTo(W2, CV_32F);
Mat projectionMatrix = (W2.t()*W1.t()).t();
//testing
#pragma omp parallel for private(temp)
for(uint i=0; i<nTestingSketches; i++){
temp = *(testingSketchesDescriptors[i]);
temp = bag(temp, bag_indexes, 154);
temp = k.projectProbe(temp);
temp = projectionMatrix.t()*(temp-meanX);
if(b==0){
testingSketchesDescriptorsBag[i] = new Mat();
*(testingSketchesDescriptorsBag[i]) = temp.clone();
}
else{
vconcat(*(testingSketchesDescriptorsBag[i]), temp, *(testingSketchesDescriptorsBag[i]));
}
}
#pragma omp parallel for private(temp)
for(uint i=0; i<nTestingPhotos; i++){
temp = *(testingPhotosDescriptors[i]);
temp = bag(temp, bag_indexes, 154);
temp = k.projectGallery(temp);
temp = projectionMatrix.t()*(temp-meanX);
if(b==0){
testingPhotosDescriptorsBag[i] = new Mat();
*(testingPhotosDescriptorsBag[i]) = temp.clone();
}
else{
vconcat(*(testingPhotosDescriptorsBag[i]), temp, *(testingPhotosDescriptorsBag[i]));
}
}
}
Mat distancesCosine = Mat::zeros(nTestingSketches,nTestingPhotos,CV_64F);
#pragma omp parallel for
for(uint i=0; i<nTestingSketches; i++){
for(uint j=0; j<nTestingPhotos; j++){
distancesCosine.at<double>(i,j) = abs(1-cosineDistance(*(testingSketchesDescriptorsBag[i]),*(testingPhotosDescriptorsBag[j])));
}
}
string file1name = "kernel-prs-" + filter + descriptor + database + to_string(nTraining) + string("cosine") + to_string(count) + string(".xml");
FileStorage file1(file1name, FileStorage::WRITE);
file1 << "distanceMatrix" << distancesCosine;
file1.release();
return 0;
}
