Rahul Raina

Interest Point Detection

In order to do interest point detection, I first computed the horizontal and vertical gradients of a gaussian when it was filtered with the image by taking the image derivatives. I then took the square of those derivatives and passed it through another gaussian filter. Using the cornerness function, I get the interest points.

% example code
% Getting the gradient of the image along the x and y axis
gaussian_filter = fspecial('gaussian', [15 15], 1);
[x_gradient, y_gradient] = imgradientxy(gaussian_filter);
Ix = imfilter(image, x_gradient);
Iy = imfilter(image, y_gradient);

% Finiding interest points
har = (g_Ix2 .* g_Iy2) - (g_Ix_Iy .* g_Ix_Iy) - alpha*((g_Ix2 + g_Iy2) .* (g_Ix2 + g_Iy2));

Get Features

For this method, I cut out image patches in order to get features. At first this didn't work because the image patches were going out of bounds, but I added some edge detection in order to only get the correct range of values for patches.

for i = 1:size(x)
    single_feature = zeros(1, 128);
    x_interest = x(i, 1);
    y_interest = y(i, 1);
    x_min = x_interest - (feature_width / 2 + 1);
    x_max = x_interest + (feature_width / 2);
    y_min = y_interest - (feature_width / 2);
    y_max = y_interest + (feature_width / 2 + 1);
    if x_min > 0 && x_max <= rows && y_min > 0 && y_max <= cols
        interest_area_mag = mag(y_min:y_max, x_min:x_max);
        interest_area_direction = direction(y_min:y_max, x_min:x_max);
        for j = 0:3
            for k = 0:3
                cell_size = feature_width / 4;
                cell_direction = interest_area_direction(j * cell_size + 1:(j+1)*cell_size, k*cell_size + 1:(k+1)*cell_size);
                cell_mag = interest_area_mag(j * cell_size + 1:(j+1)*cell_size, k*cell_size + 1:(k+1)*cell_size);
                for orient = 1:8
                    cell = sum(cell_mag(cell_direction == orient));
                    single_feature(1, 8 * (j * 4 + k) + orient) = cell;
                end

            end
        end

    end
    features(i, :) = single_feature / norm(single_feature);
end

Match Features

In order to do match features, I found the ratio of the two closest features from one image to each feature in the other image. I used pdist2 to get the eucldiean distances, sorted the matrix that it returned, and used that to find the ratios from the first and second closest features.

threshold = 0.7;
feature_distances = pdist2(features1, features2, 'euclidean');
[distances_sorted, feature_indices] = sort(feature_distances, 2);
reverse_confidence = distances_sorted(:,1)./distances_sorted(:,2);
confidences = 1./reverse_confidence(reverse_confidence < threshold);

matches = zeros(size(confidences, 1), 2);
matches(:, 1) = find(reverse_confidence < threshold);
matches(:, 2) = feature_indices(reverse_confidence < threshold, 1);

Results in a table