hi all, I'm following this tutorial to learning how to write some feature descriptor for svm and other machine learning tools, and it works really fine. However, I came up a question and search for some times and ask on several websites but got no answer : is there any way to descript the transparent feature?
for example, when I train the HOG model, the background always in same color, like
this is a hand
the background would always be the same color. It would always be classified correctly when the test data's background is simple in the HOG, for example, when our test data is in simple color background, like pure green, pure blue....
svm says: ok, I think this is a hand
but since the background always have some color, the svm always wants to find something in background, if the background is too different from the training data, the svm may cannot recognize it in some situtation.
svm says: I don't know what's this. the background is too different from your training data
As we the painter, we know that the transparency always not equal to whole white background, transparency should be transparency itself
is not 
in neural network, I can collect some same foreground data with as many as possible background to get the result like " the classifier don't care about what your background is" result. but Is there way can make it like, define and train the region which is "transparency" in the data?