Support Vector Machine one vs one training

Question

i am doing image classification using Opencv svm. My classification problem is classifying objects like cars, books, etc.Once i determine the model i add the test image to the training set and than train the system using 1 vs all training method.

But this is a very computationally expensive case since I have to train the system after adding the classified frames based on its group. Later I changed it to train after adding 5 frame, than 10 frames but it seems too expensive interms of computing.

I came across a method which I dont know how it works that is One vs One Training. I need some help in implementing 1 vs 1 training so that I dont train the entire training set of different class but just the one the test image belong to.

My confusion is : How i combine the training information after training separately?

Maybe I am treating the concept incorrectly as I am new to SVM so in that case be just help me out.

Answer 1

If I understood you correctly, you succesfully trained the SVMs on pairs of classes. Now you wonder how to combine the answers of these SVMs to get the final decision.

Usually the maximum voting strategy is used for that.

If you have a SVM that was trained to distinguish between class A and class B, depending on the output, it casts a vote for either class A or class B. You do that for all the SVMs you have trained. Then the class that got the most votes wins.