Not completely buying this. He's claiming that it's better to do vector based feathering, but it's not at all clear that his approach isn't just as computationally expensive, or more. Filters do require a lot of calculations, but they are embarrassingly parallel and also you get help from the Convolution Theorem.