Set a threshold in MulticlassSVM
I am currently testing my system(hand recognition system). It is identifying unknown gestures as known gestures, I mean that gestures that are not in my training data is being identified as one. Why is it so? It is operating in real time, I do not know where the problem lies.
Did you consider rejection? If you have no rejection label, than it is normal that unknown gestures are classified as known gestures with the biggest matching chance. I guess you do not have a rejection case?
I do not know about rejection, yes you are right I don't have a rejection case. I will explain to you waht I have done, I have given different label to each gesture for example my system should recognise only 5 gestures A,B,C,D,E and I have given them label 1,2,3,4,5,respectively. Now I do not have label F so instead of informing that the gesture is not recognised it is accepting it. What should I do now?
Well your problem is that your multiclass SVM learns seperation planes between the classes in your feature space, but that means that if you supply a label F it will look for the decision that best suits your needs. You would need to add an extra label carrying the case of negative samples, seen as the rejection case. Or you should try to infer the certainty of a decision of the classification.
Maybe interesting: http://www.macs.hw.ac.uk/bmvc2006/pap...