我在半身探测器中工作，以提高普通人探测器的性能。我知道有更多的方法来处理遮挡，但这是我在学位课程结束时被要求做的事情。我的问题是我没有得到一个好的表现，更重要的是，我得到了一个模式，其中四个代表检测的矩形显示在几乎相同的位置，甚至不代表半身。
我有一组图像，其中有414张由我自己裁剪的上半身图像，用作阳性样本，以及8520张负像。所有这些都是64x64。我提取了HOG描述符如下
int i;
string imgname, index;
HOGDescriptor hog (Size(64,64), Size(16,16), Size(8,8), Size(8,8), 9, 1, -1, HOGDescriptor::L2Hys, 0.2,false, HOGDescriptor::DEFAULT_NLEVELS, false);
vector<float> pos_rec_descript;
vector<Point> locations;
size_t SizeDesc;
SizeDesc = hog.getDescriptorSize();
FileStorage fpd ("Pos_Descriptors.yml", FileStorage::WRITE);
for (i = 1; i < 415; i++) { // 2416 images in ./pos_rec
stringstream a;
a << i;
imgname = "./pos_rec3/img" + a.str();
imgname += ".png";
Mat img = imread(imgname, CV_LOAD_IMAGE_COLOR);
hog.compute(img, pos_rec_descript, Size (16,16), Size (0,0),locations);
fpd << "Descriptores" + a.str() << pos_rec_descript;
}
fpd.release();
我对阴性样本做了同样的事情。
然后，我按如下方式训练了SVM。
#define POS 414
#define NEG 8520
#define TOTAL 8934
#define DESCRIPT 1764
float trainingData[TOTAL][DESCRIPT];
int labels[TOTAL];
fstream doc;
void set_labels(){
int i;
for (i = 0; i < TOTAL; i++){
if (i < POS) {
labels[i] = 1;
}
else{
labels[i] = -1;
}
}
return;
}
int main(int, char**)
{
FileStorage fsv ("supvec.yml", FileStorage::WRITE);
FileStorage ftd ("TrainData.yml", FileStorage::WRITE);
//FileStorage flm ("Labels.yml", FileStorage::WRITE);
FileStorage fpd ("../HOG_descriptors/Pos_Descriptors.yml", FileStorage::READ);
FileStorage fnd ("../HOG_descriptors_neg/Neg_Descriptors.yml", FileStorage::READ);
set_labels();
// Set up training data
vector <float> pos_D, neg_D, train_D ;
int k = 0;
for (int i = 1; i < POS+1; i++) {
stringstream a;
a << i;
fpd["Descriptores" + a.str()] >> pos_D;
for (int j = 0; j < pos_D.size() ; j++){
train_D.push_back(pos_D[j]);
}
}
fpd.release();
for (int i = 1; i < NEG+1; i++) {
stringstream a;
a << i;
fnd["Descriptores" + a.str()] >> neg_D;
for (int j = 0; j < neg_D.size() ; j++){
train_D.push_back(neg_D[j]);
}
}
fnd.release();
for (int i = 0; i < TOTAL; i++){
for (int j = 0; j < DESCRIPT; j++){
trainingData[i][j] = train_D[k];
k++;
}
}
Mat trainingDataMat(TOTAL, DESCRIPT, CV_32FC1, trainingData);
//memcpy(trainingDataMat.data, train_D.data(), train_D.size()*sizeof(float));
Mat labelsMat(TOTAL, 1, CV_32SC1, labels);
//ftd << "trainingDataMat" << trainingDataMat;
//flm << "labelsMat" << labelsMat;
// Train the SVM
Ptr<SVM> svm = SVM::create();
svm->setType(SVM::C_SVC);
svm->setKernel(SVM::LINEAR);
svm->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER, 100, 1e-6));
/*Ptr<TrainData> autoTrainData = TrainData::create(trainingDataMat, ROW_SAMPLE, labelsMat);
ParamGrid Cgrid = SVM::getDefaultGrid(SVM::C);
ParamGrid gammaGrid = SVM::getDefaultGrid(SVM::GAMMA);
ParamGrid pGrid = SVM::getDefaultGrid(SVM::P);
pGrid.logStep = 1;
ParamGrid nuGrid = SVM::getDefaultGrid(SVM::NU);
nuGrid.logStep = 1;
ParamGrid coeffGrid = SVM::getDefaultGrid(SVM::COEF);
coeffGrid.logStep = 1;
ParamGrid degreeGrid = SVM::getDefaultGrid(SVM::DEGREE);
degreeGrid.logStep = 1; */
cout << "Está entrenando..." << endl;
//svm->trainAuto(autoTrainData, 10, Cgrid, gammaGrid, pGrid, nuGrid, coeffGrid, degreeGrid, false);
svm->train(trainingDataMat, ROW_SAMPLE, labelsMat);
svm->save("SVM3_WS16_P0_LINEAR.yml");
我已尝试使用LINEAR和RBF内核（这就是为什么你可以看到代码的自动调整部分，我曾经在SVM类型之间进行交换），但它们似乎都不起作用。实际上，它们给出了几乎相同的响应，这让我觉得，可能是训练阶段或检测阶段（下面的代码）正在破坏整个项目。
这是我为HOG探测器加载SVM并尝试覆盖图像的方法
using namespace cv;
using namespace std;
using namespace cv::ml;
// static void help()
// {
// printf(
// "\nDemonstrate the use of the HoG descriptor using\n"
// " HOGDescriptor::hog.setSVMDetector(HOGDescriptor::getDefaultPeopleDetector());\n"
// "Usage:\n"
// "./peopledetect (<image_filename> | <image_list>.txt)\n\n");
// }
void get_svm_detector(const Ptr<SVM>& svm, vector< float > & hog_detector );
void get_svm_detector(const Ptr<SVM>& svm, vector< float > & hog_detector )
{
// get the support vectors
Mat sv = svm->getSupportVectors();
const int sv_total = sv.rows;
// get the decision function
Mat alpha, svidx;
double rho = svm->getDecisionFunction(0, alpha, svidx);
CV_Assert( alpha.total() == 1 && svidx.total() == 1 && sv_total == 1 );
CV_Assert( (alpha.type() == CV_64F && alpha.at<double>(0) == 1.) ||
(alpha.type() == CV_32F && alpha.at<float>(0) == 1.f) );
CV_Assert( sv.type() == CV_32F );
hog_detector.clear();
hog_detector.resize(sv.cols + 1);
memcpy(&hog_detector[0], sv.ptr(), sv.cols*sizeof(hog_detector[0]));
hog_detector[sv.cols] = (float)-rho;
}
int main(int argc, char** argv)
{
Mat img;
FILE* f = 0;
char _filename[1024];
if( argc == 1 )
{
printf("Usage: peopledetect (People_imgs | People_imgs.txt)\n");
return 0;
}
img = imread(argv[1]);
if( img.data )
{
strcpy(_filename, argv[1]);
}
else
{
f = fopen(argv[1], "rt");
if(!f)
{
fprintf( stderr, "ERROR: the specified file could not be loaded\n");
return -1;
}
}
// Load SVM
Ptr<SVM> svm = SVM::create();
svm = cv::Algorithm::load<ml::SVM>("../SVM_Train/SVM3_WS16_P0_LINEAR.yml");
HOGDescriptor hog (Size(64,64), Size(16,16), Size(8,8), Size(8,8), 9, 1, -1, HOGDescriptor::L2Hys, 0.2,false, HOGDescriptor::DEFAULT_NLEVELS, false);
vector <float> hog_detector;
get_svm_detector (svm, hog_detector);
hog.setSVMDetector(hog_detector);
namedWindow("people detector", 1);
for(;;)
{
char* filename = _filename;
if(f)
{
if(!fgets(filename, (int)sizeof(_filename)-2, f))
break;
//while(*filename && isspace(*filename))
// ++filename;
if(filename[0] == '#')
continue;
int l = (int)strlen(filename);
while(l > 0 && isspace(filename[l-1]))
--l;
filename[l] = '\0';
img = imread(filename);
}
printf("%s:\n", filename);
if(!img.data)
continue;
fflush(stdout);
vector<Rect> found, found_filtered, searchLocations;
vector<double> found_weights;
double t = (double)getTickCount();
// run the detector with default parameters. to get a higher hit-rate
// (and more false alarms, respectively), decrease the hitThreshold and
// groupThreshold (set groupThreshold to 0 to turn off the grouping completely).
hog.detectMultiScale(img, found, found_weights, 0, Size(16,16), Size(0,0), 1.01, 2);
//hog.detect(img, found, 0, Size(16,16), Size(0,0), searchLocations);
t = (double)getTickCount() - t;
printf("tdetection time = %gms\n", t*1000./cv::getTickFrequency());
size_t i, j;
for( i = 0; i < found.size(); i++ )
{
Rect r = found[i];
for( j = 0; j < found.size(); j++ )
if( j != i && (r & found[j]) == r)
break;
if( j == found.size() )
found_filtered.push_back(r);
}
for( i = 0; i < found_filtered.size(); i++ )
{
Rect r = found_filtered[i];
// the HOG detector returns slightly larger rectangles than the real objects.
// so we slightly shrink the rectangles to get a nicer output.
r.x += cvRound(r.width*0.1);
r.width = cvRound(r.width*0.7);
r.y += cvRound(r.height*0.07);
r.height = cvRound(r.height*0.7);
rectangle(img, r.tl(), r.br(), cv::Scalar(0,255,0), 2);
imshow("people detector", img);
waitKey(0);
}
//imshow("people detector", img);
//string imgname = "./Responses/Win_Stride16_4.png";
//imwrite(imgname, img);
int c = waitKey(0) & 255;
if( c == 'q' || c == 'Q' || !f)
break;
}
if(f)
fclose(f);
return 0;
}
我已经检查了描述符的所有维度，每个Mat似乎都没问题。但是当我使用detectMultiScale时，它会显示如下内容：
Image 1: It's strange because is missing lots of detections
Image 2: Here I realized there was a kind of pattern with this 4 rects
我的问题是无论我改变什么（在detectMultiScale中使用描述符，内核，winStride和Padding），总会有非常相似的响应，没有任何迹象表明那里有正确的检测。
我不太确定我是如何向HOG提供支持向量的，但这是我发现它的唯一方法（在StackOverflow的帖子中找到它）。
如果你们中的任何人知道这里发生了什么，以及为什么回复没有从一种配置改变到另一种配置，我将非常感谢。这段代码让我头疼几周以来。我一直在改变功能上的参数，改变HOG，改变内核，尝试不同的图像集，但似乎没有任何改变最终结果。
