using OpenCV 3.1, I'm using the SVM learning algorithm with a non-linear kernel.
I implemented a function:
void extractFeatures(cv::Mat& img, std::vector<double>& meta, std::vector<double>& features)
which takes certain information from img and meta, manipulates it and stores it in features which I then give to SVM's train(...) or predict(...).
My training data has all the meta information, but in my use case the images I want to predict on might not have.
I thought of two solutions:
- draw values uniformly at random for the missing features (obviously bad for predicting but at least no bias).
- zero out these features so they don't contribute to the prediction at all.
I would prefer solution 2, but I've run into a problem zeroing out features in the kerneled space (when using a non-linear kernel).
I woud appreciate a way to implement solution 2 or something close. Also, if someone has other recomendations for dealing with missing data I'm open to suggestions.