In order to make the descriptor rotation invariance, a
reference direction for each keypoint is assigned using local
features of an image. The magnitude and direction for each
keypoint detected by the DOG pyramid are computed using
Equation (4) and Equation (5) separately.