# Why SVM with RBF kernel always classify to one group?

I've downloaded Dog vs Cat from kaggle dataset and utilize OpenCv 3.2 Machine Learning library and c++ language,

And I choose 60-40(percent for train/test) from training set(kaggle test set do not have labels),I did grayscale and resize all images to 100x100 pixel and i calculate HOG for each one of them and I put all HOG features into the big matrix(each row for one image),

And I choose Support Vector Machine with RBF Kernel and with Gamma = 0.10625(I've tested 1.0-0.0001 for Gamma) and C = 1.5 (I've tested 1-15 for C).

The problem is when i finnaly use svm->predict(testMat, testResponse); (testMat is matrix with test HOG features and all perdiction goes to testResponse) , it always retun 0 for all test samples or 1 for all if remove some of the train set images.