SIFT feature

Other reference
In my continuous struggle to remember things that I once learned, this part I will deal with SIFT feature.
SIFT proposed by David Lowe-extract some local features in an image that are invariant to

• Scale
• Illumination
• Rotation
• Affine transformation (of the form Ax+T, which is another thing I always forget.)

The theory behind the SIFT requires more understanding, so I will leave it for later. Here I will deal with the outcome of the SIFT. Given an image, we can compute several key-points in the image along with its 128 bin long histogram of descriptor. For example given an image of building I can get following SIFT descriptors using VL_FEAT implementation of SIFT:

f: [4x683 double]
d: [128x683 uint8]
name: 'b10_1.jpg'
size: 683

The there are 683 frames found for the image 'b10_1.jpg'. For each of the 683 frames we also have a 128-Dimensional descriptor. The first frame, for example, has following attibutes:
trainSet(1).f(:,4)
x_location_of_the_disc = 12.9428  % in the column direction in the image
y_location_of_the_disc = 10.3563  % in the row direction  
scale = 2.0881
orientation = -0.0163
If the frames are plotted using vl_feat is function then it will look something like this (it is a different image though):

The yellow discs denote some random frames.