Dense sift using VLFeat

Question

I am having trouble understanding the dense sift documentation in vlfeat shown at http://www.vlfeat.org/matlab/vl_dsift.html. They have mentioned a bin size parameter for the width of the grid at a pixel and have mentioned a step parameter for vertical and horizontal displacement.

I test these parameters using the following code

[f, d] = vl_dsift(single(image), 'size', 16 , 'step' , 5 ) ;

where my image is a 50X50 matrix. I expect the keyframes to start from (1,1) move to (1,6) #displacement of 5 horizontally and so on.

But the result that I get has only 1 sift descriptor at (25,25) point.

Does anyone know what these parameters actually mean and if I am applying them correctly ? Is there any way to apply it in the above mentioned pattern (starting from top left with vertical and horizontal displacement of step size apart from manually getting the locations)

Answer 1

They wrote

The scale of the extracted descriptors is controlled by the option SIZE, i.e. the width in pixels of a spatial bin (recall that a SIFT descriptor is a spatial histogram with 4 x 4 bins).

So, it computes SIFT of the square part of the image with size 4*16 = 64 x 64 pixels. Your whole image is less than needed area for computation. I`m wondering, why you get even that one. Also, even if it would be big enough, you cannot calculate SIFT of the point (1,1), because there is only south-east part of the patch is defined, others are outside the image.