Siqi Ma (dollymsq)

Feature matching of Notre Dame.

Algorithm

We used a simplified version of SIFT pipiline here, it mainly consists of 3 steps.

1. Feature point detection;
2. Feature description;
3. Feature matching.

Feature point detection

I just simply follow the instructions in the handout, and made several little adjustments to make the project run faster.

1. Instead of calculating gradient directly, I used horizontal edge filter and vertical edge filter to get the Ix, Iy matrix of the original picture. This method is much more efficient.
2. I normalized the harrison matrix to make it easier to determine the threshold value for the binary picture.
3. I used BWCONNCOMP and COLFILT to get a local maximum of cornerness. Before doing the non-maximum suppression, it will get thousands of feature points and it's not very efficient. After non-maximum suppression, I got only about one tenth the amount of points from each image and got half the amount matched, and almost 65% matches of that are correct. This is very computationally efficient in all three steps.
   

   Image after tackled by BWCONNCOMP and COLFILT, then get a local maximum point in every white patch.

4. In order to suppress the effect of picutre edge. I pad the array within the image to simply make it black.

        for i=1:cutoff_frequency*2+4
            BW(i, :) = 0;
            BW(size(image,1)-i+1,:) = 0;
            BW(:,i) = 0;
            BW(:, size(image,2)-i+1) = 0;
        end
    

Feature description

Several adjustments:

1. To make the gradient more accurate and noise eliminated, I filtered the image first.
2. In order to make computation more efficient, instead of calculating gradient directly I used horizontal edge filter and vertical edge filter to get the Ix, Iy matrix of the original picture.
3. I used the function atan() to get the angle of the gradient. As the atan() is in the range of [-PI/2, PI/2], I used a little trick here. And I generate a histogram by dividing the degree 2*PI to eight bins. Every point with a gradient degree just fall into the closest bin. Say, if the degree of a point is PI/6 and the magnitute is 1.5, then it contributes to the (0, PI/4) by 1.5
4. I have done the "normalize -> threshold -> normalize again" step. It has improved performance of the project from 50% good matches to 60% good matches.
5. The feature-width is set 52 here to get a more accurate and sufficient description of the feature.

Feature matching

Several adjustments:

1. Set the NNDR ratio as 0.75

Results in a table

Nore Dame Got 237 feature points from left image, 319 feature points from right image; 71 matches in the end, 50 matches are correct. (70% good matches)

House Got 41 feature points from left image, 43 feature points from right image; 20 matches in the end, 15 matches are correct judged by my eyes. (75% good matches)