I have followed many tutorials on SIFT and have a good understanding of how key points are detected.But while extracting SIFT feature vectors where does the keypoint scale play the role.
As I understood, we just take a 16 X 16 window around a keypoint and store magnitudes and orientation into the bins.
I could understand where the keypoint orientation is used though not the scale.
SIFT takes window [scale X scale] * coef (usually coef = 3*sqrt(3)) around keypoint, maps it into 41x41 (fixed) patch and than does mag&ori stuff on it. Patch can be 16x16, but with poorer results .
It was a little difficult to me to understand your question, maybe you can be more direct? Anyways, here are some insights on SIFT:
The scale should be taken into account in the feature extraction that is performed in the neighbourhood. Usually, this is done by setting a Gaussian function around the keypoint with a variance proportional to the scale in which the point was detected. This function serves as weights for the estimation of the histograms of gradients.
Also, when you refer to keypoint orientation I think what you are referring to is to the most usual orientation in the neighbourhood of that point. This is computed by looking for the largest bin in the histogram and this orientation is stored in order to have rotation invariance for the points.
I hope that helps, cheers.
