I am doing a SVM classification problem on Matlab. My features are HOG features (length = 4356).
My procedure is as follows.
1.extract 200 positive windows and 200 negative windows. 2.extract HOG features of the above samples. 2. scale the features and remove the NaN features. (this gives 2904 features) 3. Do a grid search and 3-fold cross validation to find c and g values 4. train the whole training set using the best c and g 5. extract the test data set hog features and scale it with same parameters as training set. 5. test the test data set
I have a large test data set (3000 samples) to test and I already know that inside the test samples I should have only few (~10) positive samples.
However, above process gives me a 0% prediction accuracy. In fact, none of the positive samples are detected. All are detected as negatives.
Where am I doing wrong? Will having a small no of training samples (400) and large no of testing samples (3000) be a problem?
