Kamal Shadi 902922877

CS 4495 / 6476 Project 2: Local Feature Matching

Part 1 - Interest point detection

I first blure the image using:

gauss_window= fspecial('Gaussian', 5, 1);
Bimage = imfilter(image,gauss_window);

The blurred image is used to calculate the image xy gradient using matlab imgradientxy function. Boundary elements of matrix R (corner response matrix) is zeroed out.

non-maximum suppression was done using matlab colfit function.

Part 2 - Local feature description

I used what we discussed in the class. This was an easier step. The guassia filter with follwoing characteristics:

Variance = 1

Window size = 16*16

 

was used.

Please look out the code below for gradient and summation calulation used for local feature detection.

for i = 1:length(x)
    Y_part = mod(y(i)-7:y(i)+8,size(Bimage,1)); Y_part(Y_part == 0) = size(Bimage,1); 
    X_part = mod(x(i)-7:x(i)+8,size(Bimage,2)); X_part(X_part == 0) = size(Bimage,2);
    window = image(Y_part,X_part);
    gauss_filter = fspecial('Gaussian', 3, 1);
    blur_window = imfilter(window,gauss_filter,'symmetric');
   [Gmag,Gdir] = imgradient(blur_window);
   Gdir(Gdir < 0) = 360 + Gdir(Gdir < 0);
   Guass_weighted = fspecial('Gaussian', feature_width, feature_width/2);
   Gmag = Guass_weighted.*Gmag;
   count = 1;
   for k = 1:size(orientations,2)-1
       for m = 0:3
           for j = 0:3
            g_angle = Gdir(m*4+1:m*4+4,j*4+1:j*4+4);
            g_mag = Gmag(m*4+1:m*4+4,j*4+1:j*4+4);
            idx = g_angle >= orientations(k) & g_angle < orientations(k+1);
            features(i,count) = sum(g_mag(idx));
            count = count+1;
           end
       end
   end
end

 

Part 3 - Feature matching

I simplt used pairwise euclidean distance for feature matching. Ration test was also applied first to second nearest neighbor and the results were thresholded on the confidence.

Results and Discussion

Here is my result line for all three images:

• Notre Dame: 79 total good matches, 21 total bad matches. 0.79% accuracy
• Mount Rushmore: 39 total good matches, 61 total bad matches. 0.39% accuracy
• Gaudi: 3 total good matches, 97 total bad matches. 0.03% accuracy

The corresponding images for all three are seen below:

Notre Dame:

Good (green) and bad (red) matches

Interest point matching

Interest points

Rushmore:

Good (green) and bad (red) matches

Interest point matching

Interest points

Gaudi:

Good (green) and bad (red) matches

Interest point matching

Interest points

 

Extra Credit: PCA

I used matlab PCA to decrease the number of features to the 32 most variant features. Here is the code for my PCA:

% PCA

  features = features./repmat(sum(features,2),1,128);

  outPCA = pca(features);

  pca_points = features*outPCA;

  pca_points = pca_points(:,1:32);

  pca_points = pca_points./repmat(sum(pca_points,2),1,32);

features = pca_points;

For some reason PCA totally wrecked my pipeline and I was not able to trace the problem in the time. Hope you still consider my endavour in doing so.