source: birds/createmilbirds.m @ 62

% create a MIL dataset from the original WAV-files, segmentation of the
% spectrograms, and computation of features on the segmented regions

% some settings:
windowlen = 512;
fmax = 256;
intens_thr = 0.8;   % remove 80% of the signal??
f_min = 2000;       % frequency threshold (everything below is removed)

% load the 'meta' data like labels and filenames
dpath = fullfile(mildatapath,'birds');
[bagid,names] = textread(fullfile(dpath,'id2filename.txt'),'%n%s','headerlines',1);
[bagid2,labstr] = textread(fullfile(dpath,'hja_birdsong_bag_labels.txt'),'%n%s','headerlines',1);

if any(bagid~=bagid2)
        error('Bagid''s do not match.');
end
G = fspecial('gaussian',[5 5],2);

% run over the files, and get the features:
B = size(bagid,1);
x = cell(B,1);
baglab = zeros(B,13);
instlab = '';
bagid = [];
for i=1:B
        %load the signal;
        [signal,fs] = wavread(fullfile(dpath,'wavs',names{i}(2:end)));
        [S,f,t] = spectrogram(signal,windowlen,windowlen/2,fmax,fs);
        % smooth and threshold the spectrogram:
        I = imfilter(abs(S),G,'same');
        mask = (I>dd_threshold(I(:),intens_thr));
        mask(f<f_min) = 0;
        % find interesting regions:
        props = regionprops(bwlabel(mask),abs(S));
        bloblab = bwlabel(mask);
        Nseg = max(unique(bloblab));

        % run over blobs:
        im = abs(S);
        thisx = zeros(Nseg,2);
        for j=1:Nseg
                ix = (bloblab==j);
                thisx(j,:) = [props(j).Area props(j).Centroid];
                % don't forget:
                bagid(end+1) = i;
        end
        x{i} = thisx;

        % get the labels right:
        eval(['baglab(i,[',labstr{i}(2:end),'])=1;']);
end

% create a dataset
a = genmil(x);