wolffd@0: function varargout = mirnovelty(orig,varargin)
wolffd@0: %   n = mirnovelty(m) evaluates the novelty score from a similarity matrix.
wolffd@0: %   [n,m] = mirnovelty(m) also return the similarity matrix.
wolffd@0: %   Optional argument: 
wolffd@0: %       mirnovelty(...,'Distance',f) specifies the name of a dissimilarity
wolffd@0: %           distance function, from those proposed in the Statistics Toolbox
wolffd@0: %               (help pdist).
wolffd@0: %           default value: f = 'cosine'
wolffd@0: %       mirnovelty(...,'Similarity',f) specifies the name of a similarity
wolffd@0: %           measure function. This function is applied to the result of the
wolffd@0: %           distance function. cf. mirsimatrix 
wolffd@0: %           default value: f = 'exponential'
wolffd@0: %               corresponding to f(x) = exp(-x)
wolffd@0: %       mirnovelty(...,'KernelSize',s) or more simply mirnovelty(...,s) 
wolffd@0: %           specifies the length of the gaussian kernel, in samples.
wolffd@0: %           default value: s = 64.
wolffd@0: %       mirnovelty(...,'Normal',0) does not normalize the novelty curve
wolffd@0: %           between the values 0 and 1.
wolffd@0: %
wolffd@0: %   Foote, J. & Cooper, M. (2003). Media Segmentation using Self-Similarity
wolffd@0: % Decomposition,. In Proc. SPIE Storage and Retrieval for Multimedia
wolffd@0: % Databases, Vol. 5021, pp. 167-75.
wolffd@0: 
wolffd@0:         dist.key = 'Distance';
wolffd@0:         dist.type = 'String';
wolffd@0:         dist.default = 'cosine';
wolffd@0:     option.dist = dist;
wolffd@0: 
wolffd@0:         sm.key = {'Measure','Similarity'};
wolffd@0:         sm.type = 'String';
wolffd@0:         sm.default = 'exponential';
wolffd@0:     option.sm = sm;
wolffd@0: 
wolffd@0:         K.key = {'KernelSize','Width'};
wolffd@0:         K.type = 'Integer';
wolffd@0:         K.default = 64;
wolffd@0:     option.K = K;
wolffd@0:     
wolffd@0:         transf.type = 'String';
wolffd@0:         transf.default = 'TimeLag';
wolffd@0:         transf.choice = {'Horizontal','TimeLag'};
wolffd@0:     option.transf = transf;
wolffd@0: 
wolffd@0:         normal.key = 'Normal';
wolffd@0:         normal.type = 'Boolean';
wolffd@0:         normal.default = 1;
wolffd@0:         normal.when = 'After';
wolffd@0:     option.normal = normal;
wolffd@0:     
wolffd@0: specif.option = option;
wolffd@0: specif.combineframes = @combineframes;
wolffd@0: specif.nochunk = 1;
wolffd@0: varargout = mirfunction(@mirnovelty,orig,varargin,nargout,specif,@init,@main);
wolffd@0:     
wolffd@0: 
wolffd@0: function [x type] = init(x,option)
wolffd@0: type = 'mirscalar';
wolffd@0: if not(isamir(x,'mirscalar') && strcmp(get(x,'Title'),'Novelty'))
wolffd@0:     x = mirsimatrix(x,'Distance',option.dist,'Similarity',option.sm,...
wolffd@0:                       'Width',option.K,option.transf);
wolffd@0: end
wolffd@0: if isa(x,'mirdesign')
wolffd@0:     x = set(x,'Overlap',ceil(option.K));
wolffd@0: end
wolffd@0: 
wolffd@0: 
wolffd@0: function y = main(orig,option,postoption)
wolffd@0: if iscell(orig)
wolffd@0:     orig = orig{1};
wolffd@0: end
wolffd@0: if not(isa(orig,'mirscalar'))
wolffd@0:     s = get(orig,'Data');
wolffd@0:     dw = get(orig,'DiagWidth');
wolffd@0:     for k = 1:length(s)
wolffd@0:         if isnumeric(dw)
wolffd@0:             dwk = dw;
wolffd@0:         else
wolffd@0:             dwk = dw{k};
wolffd@0:         end
wolffd@0:         if option.K
wolffd@0:             cgs = min(option.K,dwk);
wolffd@0:         else
wolffd@0:             cgs = dwk;
wolffd@0:         end
wolffd@0:         cg = checkergauss(cgs,option.transf);
wolffd@0:         disp('Computing convolution, please wait...')
wolffd@0:         for z = 1:length(s{k})
wolffd@0:             sz = s{k}{z};
wolffd@0:             szm = max(max(sz));
wolffd@0:             for i = find(isnan(sz))
wolffd@0:                 sz(i) = szm;
wolffd@0:             end
wolffd@0:             cv = convolve2(sz,cg,'same');
wolffd@0:             nl = size(cv,1);
wolffd@0:             nc = size(cv,2);
wolffd@0:             if nl == 0
wolffd@0:                 warning('WARNING IN NOVELTY: No frame decomposition. The novelty score cannot be computed.');
wolffd@0:                 score{k}{z} = [];
wolffd@0:             else
wolffd@0:                 sco = cv(floor(size(cv,1)/2),:);
wolffd@0:                 incr = find(diff(sco)>=0);
wolffd@0:                 if not(isempty(incr))
wolffd@0:                     decr = find(diff(sco)<=0);
wolffd@0:                     sco(1:incr(1)-1) = NaN(1,incr(1)-1);
wolffd@0:                     if not(isempty(decr))
wolffd@0:                         sco(decr(end)+1:end) = NaN(1,length(sco)-decr(end));
wolffd@0:                     end
wolffd@0:                     incr = find(diff(sco)>=0);
wolffd@0:                     sco2 = sco;
wolffd@0:                     if not(isempty(incr))
wolffd@0:                         sco2 = sco2(1:incr(end)+1);
wolffd@0:                     end
wolffd@0:                     decr = find(diff(sco)<=0);
wolffd@0:                     if not(isempty(decr)) && decr(1)>2
wolffd@0:                         sco2 = sco2(decr(1)-1:end);
wolffd@0:                     end
wolffd@0:                     mins = min(sco2);
wolffd@0:                     rang = find(sco>= mins);
wolffd@0:                     if not(isempty(rang))
wolffd@0:                         sco(1:rang(1)-1) = NaN(1,rang(1)-1);
wolffd@0:                         sco(rang(end)+1:end) = NaN(1,length(sco)-rang(end));
wolffd@0:                     end
wolffd@0:                 end
wolffd@0:                 score{k}{z} = sco;
wolffd@0:             end
wolffd@0:         end
wolffd@0:     end
wolffd@0: else
wolffd@0:     score = get(orig,'Data');
wolffd@0: end
wolffd@0: if not(isempty(postoption)) && postoption.normal
wolffd@0:     for k = 1:length(score)
wolffd@0:         for l = 1:length(score{k})
wolffd@0:             sco = score{k}{l};
wolffd@0:             sco = (sco-min(sco))/(max(sco)-min(sco));
wolffd@0:             score{k}{l} = sco;
wolffd@0:         end
wolffd@0:     end
wolffd@0: end
wolffd@0: n = mirscalar(orig,'Data',score,'Title','Novelty'); 
wolffd@0: y = {n orig};
wolffd@0: 
wolffd@0: 
wolffd@0: function old = combineframes(old,new)
wolffd@0: if not(iscell(old))
wolffd@0:     old = {old};
wolffd@0: end
wolffd@0: if not(iscell(new))
wolffd@0:     new = {new};
wolffd@0: end
wolffd@0: for var = 1:length(new)
wolffd@0:     ov = old{var};
wolffd@0:     nv = new{var};
wolffd@0:     ofp = get(ov,'FramePos');
wolffd@0:     ofp = ofp{1}{1};
wolffd@0:     nfp = get(nv,'FramePos');
wolffd@0:     nfp = nfp{1}{1};
wolffd@0:     od = get(ov,'Data');
wolffd@0:     od = od{1}{1};
wolffd@0:     onan = find(isnan(od));
wolffd@0:     od(onan) = [];
wolffd@0:     ofp(:,onan) = [];
wolffd@0:     nd = get(nv,'Data');
wolffd@0:     nd = nd{1}{1};
wolffd@0:     nnan = find(isnan(nd));
wolffd@0:     nd(nnan) = [];
wolffd@0:     nfp(:,nnan) = [];
wolffd@0:     [unused omatch nmatch] = intersect(ofp(1,:),nfp(1,:));
wolffd@0:     if isempty(omatch)
wolffd@0:         ov = set(ov,'FramePos',{{[ofp nfp]}},'Data',{{[od nd]}});
wolffd@0:     else
wolffd@0:         lm = length(omatch);
wolffd@0:         ov = set(ov,'FramePos',{{[ofp(:,1:omatch(1)-1) nfp]}},...
wolffd@0:             'Data',{{[od(1:omatch(1)-1),...
wolffd@0:                       (od(omatch).*(lm:-1:1) + nd(nmatch).*(1:lm))/(lm+1),...
wolffd@0:                       nd(nmatch(end)+1:end)]}});
wolffd@0:     end
wolffd@0:     old{var} = ov;
wolffd@0: end
wolffd@0: 
wolffd@0: 
wolffd@0: function y = checkergauss(N,transf)
wolffd@0: hN = ceil(N/2);
wolffd@0: if strcmpi(transf,'TimeLag')
wolffd@0:     y = zeros(hN,N);
wolffd@0:     for j = 1:N
wolffd@0:         for i = 1:hN
wolffd@0:             g = exp(-((i/hN)^2 + (((j-hN)/hN)^2))*4);
wolffd@0:             if j>hN && j<hN+i
wolffd@0:                 y(hN-i+1,j) = -g;
wolffd@0:             else
wolffd@0:                 y(hN-i+1,j) = g;
wolffd@0:             end
wolffd@0:         end
wolffd@0:     end
wolffd@0: else
wolffd@0:     y = zeros(N);
wolffd@0:     for i = 1:N
wolffd@0:         for j = 1:N
wolffd@0:             g = exp(-(((i-hN)/hN)^2 + (((j-hN)/hN)^2))*4);
wolffd@0:             if xor(j>i,j>N-i)
wolffd@0:                 y(i,j) = -g;
wolffd@0:             else
wolffd@0:                 y(i,j) = g;
wolffd@0:             end
wolffd@0:         end
wolffd@0:     end
wolffd@0: end