1) For each key point, I took a 5x5 neighborhood in the detected key point scale and computed magnitude and orientation of every point in 5x5 neighborhood.Now I used the 36 bin histogram and assigned the peak orientation as the orientation of the key point.
I've just stored the peak orientations of key points in an array. How will this array useful in the later stages ?
2) Now, for every key point I took a 16x16 neighborhood and sub-divided into 16 4x4 grids. For each grid I created a 8 bin histogram , creating a 128 descriptor for each key point.
How will this descriptor is robust to orientation ?Please explain this point with some intuition.
Are the above steps correct ? Did I miss any step in between ?
I've computed the same steps for the 2nd image.
Now, Can anyone explain how to match the key points ? with some intuition behind .
I started building scale space with the staring sigma value as 0.7832 .
0.7832 0.7832*sqrt(2) 0.7832*2 0.7832*2*sqrt(2) Octave 1
0.7832*sqrt(2) 0.7832*2 0.7832*2*sqrt(2) 0.7832*4 Octave 2
0.7832*2 0.7832*2*sqrt(2) 0.7832*4 0.7832*4*sqrt(2) octave 3
0.7832*2*sqrt(2) 0.7832*4 0.7832*4*sqrt(2) 0.7832*8 octave 4
Then I computed the DoG . Let a DoG be d1 (difference of 0.7832*sqrt(2) scale and 0.7832*2 scale). d1 is the middle DoG for the 1st octave. Now is it sufficient to check the points only in d1 with the scale above and the below and with respect to its neighbors in order to consider it as a key point ? i.e If a key point is detected it must be in the middle of the levels of DoG right ? Now if some point is a key point then what would be the scale of the key point ?(i.e we detect the key point in the middle level DoG and it is the difference of the 2 different scales. How can we assign a scale to the DoG ? what would be the scale of the key point detected in that particular DoG?)
Our paper may help you
https://www.researchgate.net/publication/304019378_FEATURE_EXTRACTION_AND_MATCHING_USING_SCALE_INVARIENT_FEATURE_TRANSFORM_FOR_INDIAN_SIGN_LANGUAGE
Article FEATURE EXTRACTION AND MATCHING USING SCALE INVARIENT FEATUR...
- Badges
- Science topic
- Similar topics
- Mathematics
- Algebra
- Matrix