SVM training issue for a simple dataset (Opencv 2.4.9) - c++

Question

I'm trying a simple example to learn SVM in OpenCV, I'm not getting the right support vectors after training. Need some help in understanding the issue.
My code is :
#include <iostream>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/ml/ml.hpp>
using namespace cv;
using namespace std;
int main() {
Mat frame(Size(640,360), CV_8UC3, Scalar::all(255));
float train[15][2] = { {296, 296}, {296, 312}, {312, 8}, {312, 56}, {312, 88}, {328, 88}, {328, 104}, {328, 264}, {344, 8}, {344, 40}, {360, 8}, {360, 56}, {376, 8}, {376, 40}, {376, 56} };
Mat trainingDataMat(15, 2, CV_32FC1, train);
float labels[15] = { -1, -1, 1, 1, 1, 1, 1, -1, 1, 1, 1, 1, 1, 1, 1 };
Mat labelsMat(15, 1, CV_32FC1, labels);
CvSVMParams param;
param.svm_type = CvSVM::C_SVC;
param.C = 0.1;
param.kernel_type = SVM::LINEAR;
param.term_crit = TermCriteria(CV_TERMCRIT_ITER, 1000, 1e-6);
CvSVM SVM;
SVM.train(trainingDataMat, labelsMat, Mat(), Mat(), param);
cout<< "Training Finished..." << endl;
for(int i = 0; i < frame.rows; ++i) {
for(int j = 0; j < frame.cols; ++j) {
Mat sampleMat = (Mat_<float>(1,2) << i,j);
float response = SVM.predict(sampleMat);
//cout << response << endl;
if(response == 1) {
frame.at<Vec3b>(i,j)[2] = 0;
} else {
frame.at<Vec3b>(i,j)[0] = 0;
}
}
}
for(int dis = 0; dis < trainingDataMat.rows; dis++) {
if(labels[dis] == 1) {
circle(frame, Point((int)train[dis][0], (int)train[dis][1]), 3, Scalar (0, 0, 0), -1);
} else {
circle(frame, Point((int)train[dis][0], (int)train[dis][1]), 3, Scalar (0, 255, 0), -1);
}
}
int n = SVM.get_support_vector_count();
for(int i = 0; i < n; i++) {
const float* v = SVM.get_support_vector(i);
cout << "support Vectors : " << v[0] << " " << v[1] <<endl;
circle(frame,Point((int)v[0], (int)v[1]), 6, Scalar(128, 128, 128), 2, 8);
}
imwrite("frame.jpg",frame);
imshow("output", frame);
waitKey(0);
return 0;
}
Output image is attached
The SVM line is not separating the two classes as I expect.
Result for Support Vector is
support Vectors : 0 0.0125

Answer 1

The SVM should be OK. I think the problem lies in your display. When you call your circle(frame, Point((int)train[dis][0], (int)train[dis][1]), 3, Scalar (0, 0, 0), -1);, OpenCV understands that you want a circle in row number train[dis][1] and column number train[dis][0]. This is not what you want because a specificity of OpenCV is that it uses different coordinate systems for matrices and points. image.at<float>(Point(i,j)) is equivalent to image.at<float>(j,i).
Try replacing your circle calls with this:
if(labels[dis] == 1) {
circle(frame, Point((int)train[dis][1], (int)train[dis][0]), 3, Scalar (0, 0, 0), -1);
} else {
circle(frame, Point((int)train[dis][1], (int)train[dis][0]), 3, Scalar (0, 255, 0), -1);
}