What is the appropriate size of positive samples compared to negative samples?
So if I am feeding images of 80x30 to traincascade as positive samples, what size should the negative ones be? Also should they be bigger in relation to how big the object I'm trying to detect is in real life?
There is no exact answer to this question, as machine learning is a discipline very much based on trial and error, but a reasonable start is to add 2 negatives for each positive sample.