desc_test.m

clear;clc;close all;

% Image root path
root='repeatibility';
seqs = {'bark', 'bikes', 'boat', 'graf', 'leuven', 'trees', 'ubc', 'wall'};

% Image pair to consider
idx1 = '1';
% idx2 = '2';
idx2 = '4';

% Detector
% Oxford detectors: har, harlap, heslap, haraff, hesaff
% VLFeat detectors: dog, hessian, hessianlaplace, harrislaplace, multiscalehessian, multiscaleharris
% detector = 'haraff';
detector = 'hesaff';

% Descriptor(s)
% Oxford descriptors: jla, sift, gloh, mom, koen, cf, sc, spin, pca, cc
% VLFeat descriptors: sift_vl, liop, patch
% Our descriptors: pca, enc, enc_init
descs = {'sift', 'gloh', 'liop', 'pca', 'enc_init', 'enc'};
labels = {'SIFT', 'GLOH', 'LIOP', 'PCA', 'Pretrained DBN', 'DBN'};
assert(length(descs) == length(labels));

% Matching type
% nn, sim, ratio
mtype = 'sim'; % Mikolajczyk et al. used 'sim', while the LIOP folks used 'ratio'


%% Loop over image sequences
for seqc = seqs
    % Loops over cells return single cells
    seq = seqc{1};
    
    % Open a figure for current results
    figure('Name', seq)
    
    % Get extension, hacky
    ext = '.ppm';
    if strcmpi(seq, 'boat'), ext = '.pgm'; end
    
    % Allocate plot data (the 20 comes from the external evaluation script for
    % repeatibility)
    precision_nn = zeros(20, numel(descs));
    recall_nn = zeros(20, numel(descs));
    precision_sim = zeros(20, numel(descs));
    recall_sim = zeros(20, numel(descs));
    precision_ratio = zeros(20, numel(descs));
    recall_ratio = zeros(20, numel(descs));
    corresp = zeros(1, numel(descs));
    
    % Start gathering descriptor matching results
    for i = 1:numel(descs)
        desc = descs{i};
        
        % Derived file names
        img1 = [root '/' seq '/img' idx1 ext];
        img2 = [root '/' seq '/img' idx2 ext];
        H = [root '/' seq '/H' idx1 'to' idx2 'p'];
        det1 = [img1 '.' detector];
        det2 = [img2 '.' detector];
        desc1 = [det1 '.' desc];
        desc2 = [det2 '.' desc];
        
        assert(exist(img1, 'file') > 0, ['Image file ' img1 ' does not exist!']);
        assert(exist(img2, 'file') > 0, ['Image file ' img2 ' does not exist!']);
        assert(exist(det1, 'file') > 0);
        assert(exist(det2, 'file') > 0);
        assert(exist(desc1, 'file') > 0, ['File ' desc1 ' does not exist!']);
        assert(exist(desc2, 'file') > 0, ['File ' desc2 ' does not exist!']);
        
        [v_overlap,v_repeatability,v_nb_of_corespondences,matching_score,nb_of_matches,twi] = repeatability(...
            desc1,...
            desc2,...
            H,...
            img1,...
            img2,...
            0);
        
        [correct_match_nn,total_match_nn,correct_match_sim,total_match_sim,correct_match_rn,total_match_rn] = descperf(...
            desc1,...
            desc2,...
            H,...
            img1,...
            img2,...
            v_nb_of_corespondences(5),...
            twi);
        
        % Get precision recall
        if strcmpi(mtype, 'nn')
            corresp(i) = v_nb_of_corespondences(5);
            recall_nn(:,i) = correct_match_nn ./ corresp(i);
            precision_nn(:,i) = (total_match_nn - correct_match_nn) ./ total_match_nn;
        elseif strcmpi(mtype, 'sim')
            corresp(i) = sum(sum(twi));
            recall_sim(:,i) = correct_match_sim ./ corresp(i);
            precision_sim(:,i) = (total_match_sim - correct_match_sim) ./ total_match_sim;
        elseif  strcmpi(mtype, 'ratio')
            corresp(i) = v_nb_of_corespondences(5);
            % This has potentially fewer samples than 20
            Nratio = size(correct_match_rn, 2);
            if Nratio == 20
                recall_ratio(:,i) = correct_match_rn ./ corresp(i);
                precision_ratio(:,i) = (total_match_rn - correct_match_rn) ./ total_match_rn;
            elseif Nratio < 20
                recall_ratio(1:Nratio,i) = correct_match_rn ./ corresp(i);
                precision_ratio(1:Nratio,i) = (total_match_rn - correct_match_rn) ./ total_match_rn;
                recall_ratio(Nratio+1:end,:) = nan;
                precision_ratio(Nratio+1:end,:) = nan;
            else
                error('Invalid number of ratio PR samples!');
            end
        else
            error('No such matching strategy: %s!', mtype);
        end
    end
    
    % Plot
    if strcmpi(mtype, 'nn')
        plot(precision_nn, recall_nn, 'LineWidth', 1.5);
        maxf1 = max(2*(1-precision_nn).*recall_nn./((1-precision_nn)+recall_nn), [], 1);
    elseif strcmpi(mtype, 'sim')
        plot(precision_sim, recall_sim, 'LineWidth', 1.5);
        maxf1 = max(2*(1-precision_sim).*recall_sim./((1-precision_sim)+recall_sim), [], 1);
    elseif strcmpi(mtype, 'ratio')
        plot(precision_ratio, recall_ratio, 'LineWidth', 1.5);
        maxf1 = max(2*(1-precision_ratio).*recall_ratio./((1-precision_ratio)+recall_ratio), [], 1);
    else
        error('No such matching strategy: %s!', mtype);
    end
    legends = labels;
    for i = 1:length(legends), legends{i} = [legends{i} ' ' sprintf('(%.4f)', maxf1(i))]; end
    legend(legends, 'Interpreter', 'none', 'Location', 'NorthWest')
    title([seq ' ' idx1 '-' idx2 ' (' detector ', ' mtype ', ' num2str(corresp(end)) ' corr)'])
    xlabel('1-Precision');ylabel('Recall');
    xlim([0 1]);ylim([0 1]);
    ax=gca;ax.XTick=0:0.1:1;ax.YTick=0:0.1:1;
    grid;
    
    % Print some overall results
    fprintf('Max F1 scores (%s matching)\n', mtype);
    for i = 1:numel(labels), fprintf('\t%s: %.6f\n', labels{i}, maxf1(i)); end
    
    drawnow
    
    savefig([detector '_' seq '_' idx1 '-' idx2]);
end

disp 'Done!'