Question on SIFT

[freshnewegg]

Question on SIFT

[faturita]

VlFeat has many tweaks and modifications, many for performance reasons, and the SIFT calculation is not exactly the same as the one proposed by Lowe. For instance, the patch size used to calculate the final histogram of gradients does not match what is being shown #190.

In this case the real patch size used to determine the patch orientation will be:

Z = 2 * W = 2 * ( 3 * sigmaw ) = 2 * (3 * winf * sigma ) = 3 * 2*winf * sigma = 3 * 3 * sigma

This is because it goes from -W to W. Maybe someone from the developing team could clarify this further.

[freshnewegg]

hmm an optimization makes sense

int W = VL_MAX(floor (3.0 * sigmaw), 1) ;

suggests we either use the max of 1.5 * sigma or floor(3 * 1.5 * sigma) depending on how big sigma is. This seems like to say the more "uncertain" we are the larger surface area we consider.