removed some redundant files, testMIL left to be updated

computeHistogramFeatures.m

@@ -1,18 +1,20 @@
-function [features] = computeHistogramFeatures(im,fine,usechi2,smooth,ratio)
+function features = computeHistogramFeatures(im)
 % Extracts color,brighteness and texture features based on chi-square
 % histogram differences.
 % 
 %   [features] = computeHistogramFeatures(im,fine,b,usechi2,smooth,ratio)
 % 
-% Stavros Tsogkas <stavros.tsogkas@ecp.fr>
-% Last update: August 2015
+% Stavros Tsogkas <tsogkas@cs.toronto.edu>
+% Last update: February 2017
 
-if nargin<2, fine    = true;      end;  % compute features at finer scale
-if nargin<3, usechi2 = [1 1 1 1]; end;  % use chi-square distance or other
-if nargin<4, smooth  = 'savgol';  end;  % use savgol filtering 
-if nargin<5, ratio   = 2;         end;  % ratio between rectangle filter sides
 
-scales  = [6:2:14, 16:4:28, 32:8:48]; if fine, scales = [4, scales]; end
+opts.fine    = true;       % compute features at finer scale
+opts.usechi2 = [1 1 1 1];  % use chi-square distance or other
+opts.smooth  = 'savgol';   % use savgol filtering 
+opts.ratio   = 2;          % ratio between rectangle filter sides
+opts         = parseVarargin(opts, varargin);
+
+scales  = [6:2:14, 16:4:28, 32:8:48]; if opts.fine, scales = [4, scales]; end
 thetas  = (0:7)*pi/8;
 nScales = length(scales);
 nOrient = length(thetas);
@@ -48,7 +50,7 @@
         imrot   = imrot(pad+1-hMargin:end-pad+hMargin,pad+1-wMargin:end-pad+wMargin,:);
         for c=1:nChannels
             hgrad = computeHistogramGradient(imrot(:,:,c),nBins(c),...
-                scale,thetas(o),[h0,w0],usechi2(c),smooth,csim,ratio);            
+                scale,thetas(o),[h0,w0],opts.usechi2(c),opts.smooth,csim,ratio);            
             if size(hgrad,1)~=h || size(hgrad,2)~=w
                 hgrad = imresize(hgrad, [h,w],'bilinear');
             end
@@ -118,7 +120,6 @@
     % used to scale values into the unit interval
     abmin = -73;
     abmax = 95;
-%   gamma = 2.5; lab = RGB2Lab(im.^gamma);
     lab = applycform(im, makecform('srgb2lab'));
     lab(:,:,1) = lab(:,:,1) ./ 100;
     lab(:,:,2) = (lab(:,:,2) - abmin) ./ (abmax-abmin);
@@ -147,7 +148,6 @@
 
 % --- Get subsampling rates and pyramid level indexes ---------------------
 function [step, pyramidLevel] = assignPyramidLevel(scales)
-
 % step: step between rectangle scales in the same level of the pyramid
 % pyramidLevel: index of pyramid level according to scale
 
@@ -173,4 +173,4 @@
 % Pad images to avoid wrong symmetry responses at larger scales
 for isub = 1:4
     impyr{isub} = padarray(impyr{isub},[pad pad],'replicate');
-end
+end

computeSpectralFeature.m

@@ -1,4 +1,4 @@
-function [sPb,eigVec] = computeSpectralFeature(psym,nvec,radius,sigmaW,sigmaOE)
+function [spb,eigVec] = computeSpectralFeature(psym,nvec,radius,sigmaW,sigmaOE)
 % function [sPb,eigVec] = computeSpectralFeature(psym,nvec,radius,sigmaW,sigmaOE)
 % 
 % Calculate spectral feature. Code based on the code for Berkeley global
@@ -69,7 +69,7 @@
 ch_per  = [2 3 4 5 6 7 8 1];     % change theta sequence to [0,pi) clockwise
 
 eigVec = zeros(height,width,nvec-1);
-sPb    = zeros(height, width, norient);
+spb    = zeros(height, width, norient);
 for v = 1 : nvec
     if EigVal(v) > 0,
         eigVec(:,:,v) = vect(:,:,v)/sqrt(EigVal(v));
@@ -78,7 +78,7 @@
         for o = 1 : norient,
             theta = dtheta*o;
             f = oeFilter(sigmaOE, support, theta, deriv, hil);
-            sPb(:,:,ch_per(o)) = sPb(:,:,ch_per(o)) + abs(applyFilter(f, eigVec(:,:,v)));
+            spb(:,:,ch_per(o)) = spb(:,:,ch_per(o)) + abs(applyFilter(f, eigVec(:,:,v)));
         end
     end
 end

saveSpectralFeatures.m

@@ -1,51 +1,54 @@
-function saveSpectralFeatures(set, nvec, radius, sigma)
+function saveSpectralFeatures(model, nvec, radius, sigma)
 % Compute spectral feature for all images in one of the train, val, test
 % sets of the BSDS500 dataset.
 % 
-%   computeAllSpectralFeatures(set,nvec,radius,sigma);
+%   saveSpectralFeatures(set,nvec,radius,sigma);
 % 
-%   set:    one of {'train', 'val', 'test'}.
+%   model:  model used to compute soft symmetry responses that are fed as
+%           input to the spectral clustering algorithm.
 %   nvec:   number of eigenvectors used {default = 50}.
 %   radius: radius for intervening contour cue {default = 5}.
 %   sigma:  sigma for intervening contour cue. {default = 0.1}
 % 
-% Stavros Tsogkas <stavros.tsogkas@ecp.fr>
-% Last update: August 2013
+% Stavros Tsogkas <tsogkas@cs.toronto.edu>
+% Last update: February 2017
 
-if nargin<1, set    = 'train'; end
 if nargin<2, nvec   = 50;      end  % number of eigenvectors computed
 if nargin<3, radius = 5;       end  % radius for intervening contour cue
 if nargin<4, sigma  = 0.1;     end  % sigma for intervening contour cue
 
-% Set paths
-directories   = setupDirectories();
-switch set
-    case 'train'
-        imagePath = directories.bsds500imTrain;
-    case 'val'
-        imagePath = directories.bsds500imVal;
-    case 'test'
-        imagePath = directories.bsds500imTest;
-    otherwise
-        error('ERROR:Wrong dataset!')
+% Load model 
+if ischar(model)
+    spbModel = load(fullfile(paths.models,model)); spbModel = spbModel.spbModel;
+elseif isstruct(model)
+    spbModel = model;
+else error('Input model must be a path to a mat-file or a struct.')
 end
-s = load(fullfile(directories.models,'rpbModel_600K_color_loglikeNOR_balanced_trainval.mat'));
-w = s.rpbModel.w; 
-savePath = fullfile(directories.spectral,['spectral_' num2str(nvec)],set);
-imageList = dir(fullfile(imagePath,'*jpg'));
+
+% Create directory where spectral features will be stored
+mkdir(fullfile(path.spectral,['eigenvec' num2str(nvec)]));
+
+% Read image paths
+imageList = [dir(fullfile(paths.bsds500imTrain,'*jpg'));
+             dir(fullfile(paths.bsds500imVal,  '*jpg'));
+             dir(fullfile(paths.bsds500imTest, '*jpg'))];
+
+% Compute and store spectral features
 for i = 1:length(imageList)
     [~,iid,~] = fileparts(imageList(i).name);
-    if ~exist(fullfile(savePath,['spectral_' num2str(iid) '.mat']),'file')
-        fprintf('Extracting spectral features from %s set image %d/%d (iid = %s)\n',...
+    savePath = fullfile(path.spectral, ['eigenvec' num2str(nvec)], ['spectral_' num2str(iid) '.mat']);
+    if ~exist(savePath,'file')
+        fprintf('Extracting spectral features for image %d/%d (iid = %s)\n',...
             set, i, length(imageList), iid);
-        im = im2double(imread(fullfile(imagePath, imageList(i).name)));
-        ridges = rpb(im,'color',true,w);
-        [spectral,eigVec] = computeSpectralFeature(ridges.thin,nvec,radius,sigma);
+        img = im2double(imread(fullfile(imagePath, imageList(i).name)));
+        spb = spbMIL(img,spbModel.featureSet,true,spbModel.w);
+        [spectral,eigVec] = computeSpectralFeature(spb.thin,nvec,radius,sigma);
         spectral = single(spectral);
-%             for iv=1:size(eigVec,3)
-%                 fig = figure; imagesc(eigVec(:,:,iv)); print(fig,'-depsc2',sprintf('eigVec%d_%s.eps',iv+1,iid))
-%             end
-        save(fullfile(savePath,['spectral_' num2str(iid)]),'spectral')
+%         for iv=1:size(eigVec,3)
+%             fig = figure; imagesc(eigVec(:,:,iv)); 
+%             print(fig,'-depsc2',sprintf('eigVec%d_%s.eps',iv+1,iid))
+%         end
+        save(savePath,'spectral')
     end
 end
 

setPaths.m

@@ -26,4 +26,7 @@
 % models
 paths.models           = fullfile(paths.output,'mil','models');
 
+% path for storing spectral features
+path.spectral          = fullfile(paths.output, 'spectral');
+