Hello,
I have been training a svm classifier for a 2 class forgery detection problem with feature size of 18157 and number of samples = 6000. The svm type is c_svc with a rbf kernel. The C and Gamma parameters were varied for improving accuracy. Even svm_auto training method was tried with a reduced sample size. But the maximum accuracy obtained was around 60%.
The same features was given to a matlab ensemble [http://dde.binghamton.edu/download/ensemble/] with default parameters. The accuracy obtained was more than 80%. In order to maintain the same features, the features were dumped from opencv to a file and then read inside matlab and then fed to the ensemble trainer. Both opencv and matlab features were compared inside matlab and found to be same.The accuracy was still above 80%.
In order to remove the problem of low sample numbers compared to feature size, svm was trained on a sample size of 34,000 samples. Still the accuracy is only around 60%.
Why is there an accuracy difference of 20% between svm and matlab ensemble?
Regards
Amal