SIFT representation

We tried these SIFTs:

SIFT-Ori:

Here every direction and gradient votes with equal weight. This is not very discriminative since it cannot distinguish between the black circle on white background and between the random texture where all directions have the same probability.

SIFT-Grad:

The gradient magnitude is computed and its direction gets votes related to this magnitude. This turned out to have more discriminative power.

SIFT-GradWei:

This one is very similar to the SIFT-Grad but the gradient's magnitudes are multiplied with a coefficient. This coefficient decreases with the increasing distance from the center of square patch.

Overlapping SIFTs:

each of SIFTs above can have an overlapping square patches over itself.

These parameters are important for every experiment with SIFTs:

• number of square patches in horizontal direction ($H$)
• number of square patches in vertical direction ($V$)
• number of bins ($B$)

Will use the term SIFT topology by which we mean: number of square patches in horizontal direction, number of patches in vertical direction and number of bins per square patch. We will write this as: HxVxB or HxV if we do not focus on number of bins per square patch.