Annie Carlson (aecarlso)

get_interest_points

For get_interest points I first took the dx and dy of the image

maskX = [-1 0 1; -1 0 1; -1 0 1];   
maskY = maskX';

dx = conv2(image, maskX, 'same');      
dy = conv2(image, maskY, 'same');

Then I computed the three images corresponding to the outer products of these gradients, convolved with a larger Gaussian, and used these to compute the harris measure:

dxSquared = conv2(dx.^2, gaus, 'same'); 
dySquared = conv2(dy.^2, gaus, 'same');
dxy = conv2(dx.*dy, gaus, 'same');

% Compute the HCM	
cornMeas = (dxSquared.*dySquared - dxy.^2) - k*(dxSquared + dySquared).^2; %

Finally, I limited results to local maxima and added a bordermask:

maxi = colfilt(cornMeas,[feature_width,feature_width], 'sliding', @max);

border = zeros(size(cornMeas));
border(ceil(feature_width/2)+1:end-ceil(feature_width/2), ceil(feature_width/2)+1:end-ceil(feature_width/2)) = 1;
 
corrected = maxi & (cornMeas>thresh) & border;


    
[x,y] = find(corrected);

get_features

I believe get_features is buggy, due to my low performance, but I can't figure out where.

First, I find the gradient at each point.

[gradImX, gradImY] = gradient(image);

First, I go over each of the interest points found by the previous method, and and break the feature_width sized box around it into 4x4 equal sized boxes. for each pixel in these boxes I measure the gradient and bin the gradient into one of 8 bins, creating a histogram for each of my 16 boxes

[  for j=1:4,
        for k=1:4,
            cellsX(j,k,:,:)=gradImX(X-feature_width+(1+j)*(feature_width/4):X-feature_width+(2+j)*(feature_width/4)-1,Y-feature_width+(1+k)*(feature_width/4):Y-feature_width+(2+k)*(feature_width/4)-1);
            cellsY(j,k,:,:)=gradImY(X-feature_width+(1+j)*(feature_width/4):X-feature_width+(2+j)*(feature_width/4)-1,Y-feature_width+(1+k)*(feature_width/4):Y-feature_width+(2+k)*(feature_width/4)-1);
        end
    end
    %  (2) each cell should have a histogram of the local distribution of
    %    gradients in 8 orientations. Appending these histograms together will
    %    give you 4x4 x 8 = 128 dimensions.
    %cellsY(:,:,:,:) = max(cellsY(:,:,:,:),.0001); 
    for j=1:4,
        for k=1:4,
            vals = atan(cellsX(j,k,:,:) ./ cellsY(j,k,:,:));
            vals = ceil(mod(vals(:,:,:,:), 2*pi));
            for g=1:8,
                features(i,(j-1)*32+(k-1)*8+g) = [sum(vals(:,:) == g)];
            end
        end
    end

Finally, I normalize each feature vector.

features(i, :) = features(i, :)./norm(features(i, :));

match_features

For match_features I would find the 2 nearest neighbors. The the nearest neighbor score/ second highest nearest neighbor score is below some threshold, then I would classify the feature and the nearest neighbor as a match. Because of whatever error I had in the previous section, I was forced to make this threshold very high to get any results

Results

The cathedral I was able to get 10% accuracy on by fiddling with various thresholds. I believe my results were so poor because of an error in find_features