I need to understand how SIFT calculates the descriptors for the keypoints.
Intuitively, I understand that it takes each keypoint, calculates the gradients for each pixel in a neighborhood of the keypoint, and that's basically the descriptor for the keypoint. The paper mentions a coordination system rotation in the keypoint, I assume this is when the image is rotated, the keypoint descriptor doesn't change.
My question:
I'm following this implementation of SIFT. In the part of the calculation function, there is this cos/sin calculation:
I think this is related to the coordinate system rotation. Can you explain how the coordinate system is being rotated? Why does that have to do with the hist_width?
