int main() {
/* Step 0: Configuration of parameters
*/
const int bins = 9; // The number of bin to 9.
const int cells = 8; // The size of cell to cover 8 x 8 pixels.
const int blocks = 2; // The size of block to contain 2 x 2 cells.
const int featureSizes[3] = {15, 7, 36}; // The size of the training feature.
const int numberOfPositive = 2416; // The number of positive samples.
const int numberOfNegative = 6090; // The number of negative samples.
const int numberOfValidation = 40;
CvSVMParams params; // The parameters of SVM
params.svm_type = CvSVM::C_SVC;
params.kernel_type = CvSVM::LINEAR;
params.term_crit = cvTermCriteria(CV_TERMCRIT_ITER, 500000, 1e-10);
const float scales[] = { 0.21, 0.27, 0.3}; // Multiple scales
// const float threshold[] = {3.0, 3.5, 4.5}; // The threshold for each scales
const float threshold[] = { 1.5 ,1.55, 1.6 }; // The threshold for each scales
const bool drawBox = true; // Switch to draw the bounding boxes.
const float overlap = 0.2; // The overlap ration for NMS
/* Step 1: Extract HOGs of all training samples
*/
cout << "Step 1: Extract HOGs of all training samples..." << endl;
HOGExtractor ext(bins, cells, blocks); // Initialize a HOG extractor.
Mat trainSample;
Mat trainLabel;
char buf[6];
if (!fopen("..\\Descriptor\\traindata.yml", "r")) { // Delete the descriptor files if you want to extract new ones
Mat positiveDescriptor(numberOfPositive, featureSizes[0]*featureSizes[1]*featureSizes[2],
CV_32FC1, Scalar(0));
Mat positiveLabel = Mat::ones(numberOfPositive, 1, CV_32FC1);
Mat negativeDescriptor(numberOfNegative, featureSizes[0]*featureSizes[1]*featureSizes[2],
CV_32FC1, Scalar(0));
Mat negativeLabel = -Mat::ones(numberOfNegative, 1, CV_32FC1);
for (int i = 0; i < numberOfPositive; ++i) { // positive samples
cout << "Positive HOG descriptor: " << i+1 << "\\"
<< numberOfPositive << endl;
sprintf(buf, "%.6d", i+1);
Mat im = imread("..\\Dataset\\Train\\pos\\"+string(buf)+".png",
CV_LOAD_IMAGE_GRAYSCALE);
Mat tmp = ext.ExtractHOG(im);
Mat feature(1, featureSizes[0]*featureSizes[1]*featureSizes[2], CV_32FC1, tmp.data);
feature.copyTo(positiveDescriptor.row(i));
}
for (int i = 0; i < numberOfNegative; ++i) { // negative samples
cout << "Negative HOG descriptor: " << i+1 << "\\"
<< numberOfNegative << endl;
sprintf(buf, "%.6d", i+1);
Mat im = imread("..\\Dataset\\Train\\neg\\"+string(buf)+".png",
CV_LOAD_IMAGE_GRAYSCALE);
Mat tmp = ext.ExtractHOG(im);
Mat feature(1, featureSizes[0]*featureSizes[1]*featureSizes[2], CV_32FC1, tmp.data);
feature.copyTo(negativeDescriptor.row(i));
}
vconcat(positiveDescriptor, negativeDescriptor, trainSample);
vconcat(positiveLabel, negativeLabel, trainLabel);
FileStorage fs("..\\Descriptor\\traindata.yml", FileStorage::WRITE);
fs << "trainSample" << trainSample << "trainLabel" << trainLabel;
fs.release();
} else {
FileStorage fs("..\\Descriptor\\traindata.yml", FileStorage::READ);
fs["trainSample"] >> trainSample;
fs["trainLabel"] >> trainLabel;
fs.release();
}
cout << "done!" << endl;
/* Step 2: Train Linear Detector with SVM
*/
cout << "Step 2: Train Linear Detector with SVM..." << endl;
Detector dt;
if (!fopen("..\\Detector\\detector.yml", "rb")) { // Delete the detector file if you want to train a new one
dt.TrainDetector(trainSample, trainLabel, featureSizes, params);
Mat detector = dt.GetDetector();
Mat saveDet(1, featureSizes[0]*featureSizes[1]*featureSizes[2], CV_32FC1, detector.data);
FileStorage fs("..\\Detector\\detector.yml", FileStorage::WRITE);
fs << "detector" << saveDet;
fs.release();
} else {
FileStorage fs("..\\Detector\\detector.yml", FileStorage::READ);
Mat loadDet;
fs["detector"] >> loadDet;
Mat detector(featureSizes[0], featureSizes[1], CV_32FC(featureSizes[2]), loadDet.data);
dt.SetDetector(detector);
fs.release();
}
cout << "done!" << endl;
/* Step 3: Perform detection
*/
cout << "Step 3: Perform detection..." << endl;
PostProcessor pro;
Mat res;
for (int i = 0; i < numberOfValidation; ++i) {
cout << "Detect " << i+1 << "\\" << numberOfValidation << endl;
sprintf(buf, "%.5d", i+1);
Mat im = imread("..\\Dataset\\Validation\\"+string(buf)+".png",
CV_LOAD_IMAGE_GRAYSCALE);
// cout << "scales[" << 0 << "] = " << scales[0] << endl;
// cout << "threshold[" << 0 << "] = " << threshold[0] << endl;
Mat bbox = dt.MultiscaleDetection(im, ext, scales,
sizeof(scales)/sizeof(float), threshold);
Mat top = pro.NonMaxSup(bbox, overlap);
Mat ids = i*Mat::ones(top.size[0], 1, CV_32FC1); // image id
if (top.dims != 0) {
Mat temp;
hconcat(ids, top, temp); // attach the image id
if (res.dims == 0) {
res = temp;
} else {
vconcat(res, temp, res);
}
}
if (drawBox == true) {
im = imread("..\\Dataset\\Validation\\"+string(buf)+".png",
CV_LOAD_IMAGE_COLOR);
dt.DisplayDetection(im, top);
}
}
cout << "done!" << endl;
/* Step 4: Evaluation the performance of detection
*/
cout << "Step 4: Evaluation the performance of detection..." << endl;
vector<Mat> gt;
for (int i = 0; i < numberOfValidation; ++i) {
sprintf(buf, "%.5d", i+1);
FileStorage fs("..\\Dataset\\Validation\\"+string(buf)+".yml", FileStorage::READ);
cout << "..\\Dataset\\Validation\\" + string(buf) + ".yml" << endl;
Mat groundtruth;
fs["groundtruth"] >> groundtruth;
fs.release();
Mat detFlag = Mat::zeros(groundtruth.size[0], 1, CV_32FC1);
cout <<"GT(r,c): "<< groundtruth.rows <<", "<< groundtruth.cols << endl;
cout << "detFlag(r,c): " << detFlag.rows << ", " << detFlag.cols << endl;
hconcat(groundtruth, detFlag, groundtruth);
gt.push_back(groundtruth);
}
DetEvaluator eva;
vector<Mat> report;
float ap = eva.EvalDetection(res, gt, report);
cout << "AP: " << ap << endl;
FILE * pFile = fopen ("..\\Report\\report.txt", "wb");
fwrite (&ap, sizeof(float), 1, pFile);
fwrite (report[0].data, sizeof(float), report[0].total(), pFile);
fwrite (report[1].data, sizeof(float), report[1].total(), pFile);
fclose (pFile);
cout << "done!" << endl;
return 0;
}
