BOW and SVM better multiclass or multi SVM 2 classes?

edit

I have seen that in the example of BOW in OpenCV examples there is trained a SVM classifier for each class. I am wondering why not using a multi class SVM? Is it much more accurate the 2 classes SVM than the multi class one? If I do like this and if there are X classes, then I shall load X SVMs and classify X times an image...