In this part whether I should consider originally sampled 16 x 16 window around keypoint or I should consider newly transformed 16 x 16 window with respect to keypoint orientation ?

Following process is with respect to original 16 x 16 window :

1. First I have selected a window of 16x16 around keypoint as in part-I

2. I have divided 16 x 16 window around keypoint in 16 windows , each of 4 x 4.

3. Calculated Magnitude and orientation for each window of 4 x 4.

4. then I have assigned Bin no. of 8 bin histogram to respective orientation and magnitude.

5. then accumulated all entries of orientation according to bin no. and also calculated accumulated sum of magnitude according to bin no. (here I got orientation in degree, so i converted it into radians)

6. From above values i tried to analyze relation between orientation and magnitude.

7. e.g. I got 8 entries in bin no. '1' and 7 entries in bin no. '8'.

8. Value of magnitude of bin No. '1' is 2 and value of magnitude of bin no. 8 is 51.22 (What is the role of magnitude in calculation of vector of size 128 which is local descriptor of keypoint?) I have also calculated 4 x 4 window out of 16 x 16 Gaussian weighted circular window to multiply with 4 x 4 magnitude window. (But where to use this multiplication?)