change fixation plot color to be default and random over plots AND ch…

…ange the transition matrix image color(color gray --> colormap default)

Classif_from_HMM_descriptors_Coutrot.m

@@ -0,0 +1,87 @@
+%If you use this toolbox, please cite Coutrot et al., 
+%"Scanpath modeling and classification with Hidden Markov Models", Behavior
+%Research Methods, 2017
+% Classify tasks from Coutrot's dataset (conversational videos, 2 auditory conditions) with HMM-based gaze descriptors
+clear
+close all
+load HMM_descriptor_Coutrot %computed with exemple_Compute_HMM_descriptors.m
+all_stim=fieldnames(example_HMM_descriptor_Coutrot);
+
+%LDA 1st Eigen Vector
+%ini_stim=char(all_stim(1));
+%ini_descriptor=example_HMM_descriptor_Coutrot.(ini_stim).with_os.gaze_descriptor;
+%LDA_1st_Eigen_vect=NaN(size(ini_descriptor,2),length(all_stim));
+
+%Correct classification vector
+correct_classif=NaN(length(all_stim),1);
+
+% For each stimulus, a correct classification score is computed
+for istim=1:length(all_stim)
+
+    im_name=char(all_stim(istim));
+
+    %% Load HMM-based gaze descriptors from HMM_descriptor_Coutrot.mat
+    gaze_descriptor_ws=example_HMM_descriptor_Coutrot.(im_name).with_os.gaze_descriptor;
+    gaze_descriptor_wos=example_HMM_descriptor_Coutrot.(im_name).without_os.gaze_descriptor;
+
+    %% Normalization to zero mean and unit std
+    norm_gaze_descriptor_ws=zscore(gaze_descriptor_ws')';
+    norm_gaze_descriptor_wos=zscore(gaze_descriptor_wos')';
+
+    %% Regularization
+    all_task = [norm_gaze_descriptor_ws;norm_gaze_descriptor_wos];
+    lambda_I_all=0.00001*eye(size(all_task));
+    regul_gaze_descriptor_all = all_task - lambda_I_all.*all_task + lambda_I_all ;
+    regul_gaze_descriptor_ws=regul_gaze_descriptor_all(1:size(norm_gaze_descriptor_ws,1),:);
+    regul_gaze_descriptor_wos=regul_gaze_descriptor_all(size(norm_gaze_descriptor_ws,1)+1:...
+        size(norm_gaze_descriptor_ws,1)+size(norm_gaze_descriptor_wos,1),:);
+
+
+    %% Choose classes to classify
+    gaze_descriptors={regul_gaze_descriptor_ws, regul_gaze_descriptor_wos};
+    categoric_var={'with_os', 'without_os'};
+
+    %% Select type of classifier
+    classifier_type='LDA';
+    %  classifier_type='diagquadratic';
+    % classifier_type='mahalanobis';
+    %classifier_type='SVMBinary';
+    %classifier_type='SVMMultiClass';
+    %classifier_type='AdaBoostBinary';
+    %classifier_type='AdaBoostMultiClass';
+    % classifier_type='RVM';%Only for 2-class problems
+    % classifier_type='AdaBoost';%Only for 2-class problems
+    % classifier_type= 'RandomForest';
+
+    %% k-fold cross-validation
+  cross_validation=1;
+% % if cross_validation==1
+% %     leave-one-out
+% % else
+% %    'k'-cross_validation
+% % end
+    try
+        [lda_stats, success_rate] = classifier(categoric_var, gaze_descriptors,classifier_type,cross_validation);
+
+        %         %LDA 1st Eigen vector: absolute values and normalization
+        %         lda_stats.eigenvec(:,1)=abs(lda_stats.eigenvec(:,1));
+        %         LDA_1st_Eigen_vect(:,istim)=lda_stats.eigenvec(:,1)/sum(lda_stats.eigenvec(:,1));
+    catch
+        fprintf('stimuli %u could not be classified\n',istim)
+        success_rate=NaN;
+        manova_stats=NaN;
+    end
+
+    if ~mod(istim,1)
+        fprintf('stimuli %u success_rate %d\n',istim, success_rate)
+    end
+
+    correct_classif(istim)=success_rate;
+end
+
+fprintf('average correct classification score over %u stimuli is %3.1f %% (chance = %3.1f %%)\n',length(all_stim), 100*nanmean(correct_classif),100/length(categoric_var))
+hist(correct_classif)
+xlabel('Correct Classification Rate','FontSize',12,'FontWeight','bold')
+ylabel('Frequency','FontSize',12,'FontWeight','bold')
+
+% errorbar(squeeze(nanmean(abs(LDA_1st_Eigen_vect),2)),nanstd(abs(LDA_1st_Eigen_vect),0,2)/sqrt(length(all_stim)),'.')

Classif_from_HMM_descriptors_Koehler.m

@@ -0,0 +1,90 @@
+%If you use this toolbox, please cite Coutrot et al.,
+%"Scanpath modeling and classification with Hidden Markov Models", Behavior
+%Research Methods, 2017
+% Classify tasks from Koehler's dataset (static images, 3 tasks) with HMM-based gaze descriptors
+clear
+close all
+load HMM_descriptor_Koehler %computed with exemple_Compute_HMM_descriptors.m
+all_stim=fieldnames(example_HMM_descriptor);
+ini_stim=char(all_stim(1));
+ini_descriptor=example_HMM_descriptor.(ini_stim).freeview.gaze_descriptor;
+%LDA 1st Eigen Vector
+%LDA_1st_Eigen_vect=NaN(size(ini_descriptor,2),length(all_stim));
+%Correct classification vector
+correct_classif=NaN(length(all_stim),1);
+
+% For each stimulus, a correct classification score is computed
+for istim=1:length(all_stim)
+
+    im_name=char(all_stim(istim));
+
+    %% Load HMM-based gaze descriptors from example_HMM_descriptor.mat
+    gaze_descriptor_free=example_HMM_descriptor.(im_name).freeview.gaze_descriptor;
+    gaze_descriptor_sal=example_HMM_descriptor.(im_name).salview.gaze_descriptor;
+    gaze_descriptor_obj=example_HMM_descriptor.(im_name).objsearch.gaze_descriptor;
+
+    %% Normalization to zero mean and unit std
+    norm_gaze_descriptor_free=zscore(gaze_descriptor_free')';
+    norm_gaze_descriptor_sal=zscore(gaze_descriptor_sal')';
+    norm_gaze_descriptor_obj=zscore(gaze_descriptor_obj')';
+
+    %% Regularization
+    all_task = [norm_gaze_descriptor_free;norm_gaze_descriptor_sal;norm_gaze_descriptor_obj];
+    lambda_I_all=0.00001*eye(size(all_task));
+    regul_gaze_descriptor_all = all_task - lambda_I_all.*all_task + lambda_I_all ;
+    regul_gaze_descriptor_free=regul_gaze_descriptor_all(1:size(norm_gaze_descriptor_free,1),:);
+    regul_gaze_descriptor_sal=regul_gaze_descriptor_all(size(norm_gaze_descriptor_free,1)+1:size(norm_gaze_descriptor_free,1)+size(norm_gaze_descriptor_sal,1),:);
+    regul_gaze_descriptor_obj=regul_gaze_descriptor_all(size(norm_gaze_descriptor_free,1)+size(norm_gaze_descriptor_sal,1)+1:size(norm_gaze_descriptor_free,1)+size(norm_gaze_descriptor_sal,1)+size(norm_gaze_descriptor_obj,1),:);
+
+
+    %% Choose classes to classify
+    %     categoric_var={'free', 'sal', 'obj'};
+    %     gaze_descriptors={regul_gaze_descriptor_free, regul_gaze_descriptor_sal, regul_gaze_descriptor_obj};
+
+    gaze_descriptors={regul_gaze_descriptor_free, regul_gaze_descriptor_sal, regul_gaze_descriptor_obj};
+    categoric_var={'free', 'sal', 'obj'};
+
+    %% Select type of classifier
+    classifier_type='LDA';
+    %  classifier_type='diagquadratic';
+    % classifier_type='mahalanobis';
+    %classifier_type='SVMBinary';
+    %classifier_type='SVMMultiClass';
+    %classifier_type='AdaBoostBinary';
+    %classifier_type='AdaBoostMultiClass';
+    % classifier_type='RVM';%Only for 2-class problems
+    % classifier_type='AdaBoost';%Only for 2-class problems
+    % classifier_type= 'RandomForest';
+
+    %% k-fold cross-validation
+    cross_validation=1;
+    % % if cross_validation==1
+    % %     leave-one-out
+    % % else
+    % %    'k'-cross_validation
+    % % end
+    try
+        [lda_stats, success_rate] = classifier(categoric_var, gaze_descriptors,classifier_type,cross_validation);
+
+        %         %LDA 1st Eigen vector: absolute values and normalization
+        %         lda_stats.eigenvec(:,1)=abs(lda_stats.eigenvec(:,1));
+        %         LDA_1st_Eigen_vect(:,istim)=lda_stats.eigenvec(:,1)/sum(lda_stats.eigenvec(:,1));
+    catch
+        fprintf('stimuli %u could not be classified\n',istim)
+        success_rate=NaN;
+        manova_stats=NaN;
+    end
+
+    if ~mod(istim,10)
+        fprintf('stimuli %u success_rate %d\n',istim, success_rate)
+    end
+
+    correct_classif(istim)=success_rate;
+end
+
+fprintf('average correct classification score over %u stimuli is %3.1f %% (chance = %3.1f %%)\n',length(all_stim), 100*nanmean(correct_classif),100/length(categoric_var))
+hist(correct_classif,40)
+xlabel('Correct Classification Rate')
+ylabel('Frequency')
+
+% errorbar(squeeze(nanmean(abs(LDA_1st_Eigen_vect),2)),nanstd(abs(LDA_1st_Eigen_vect),0,2)/sqrt(length(all_stim)),'.')

Compute_HMM_descriptors_Coutrot.m

@@ -0,0 +1,100 @@
+%If you use this toolbox, please cite Coutrot et al., 
+%"Scanpath modeling and classification with Hidden Markov Models", Behavior
+%Research Methods, 2017
+
+% Compute HMM-based gaze descriptors for Koehler's dataset (static images, 3 tasks)
+clear
+close all
+addpath ./stimuli/Frames_Coutrot/ % where the stimuli are stored
+addpath(genpath('emhmm-toolbox'))
+load EyeData_Coutrot % contains eye positions from Coutrot & Guyader, JoV 2014
+
+% variational approach: automatically selects optimal state number from K = 1 to 3
+K = 1:3;
+% [priors transition_matrix_coeff state_centre state_cov]
+param_nb=max(K)+max(K)^2+max(K)*2+max(K)*2;
+%plot the HMM states on the stimuli
+isplotHMM=1;
+
+nstim=size(example_EyeData_Coutrot.with_os.x_pos,2); %number of stimuli
+for istim=1:nstim
+    fprintf('stim %u\n',istim)
+    %close all
+
+    % Read stimuli
+    im_name=['clip_' num2str(istim) '.jpg'];
+    im_name_struct=['clip__' num2str(istim)];
+    im= imread(im_name);
+
+    % Maximum number of observer per task
+    maxsub=max([size(example_EyeData_Coutrot.with_os.x_pos,1),size(example_EyeData_Coutrot.without_os.x_pos,1)]);
+    % Initialize gaze descriptor vector
+    gaze_descriptor_ws=NaN(maxsub,param_nb);
+    gaze_descriptor_wos=NaN(maxsub,param_nb);
+
+    % Loop on subjects
+    for isub=1:maxsub
+
+        %% Auditory Condition 1: With Orginal Soundtrack
+
+        % Extract current scanpath
+        scanpath_ws = extract_scanpath(example_EyeData_Coutrot,'with_os',isub,istim,K);
+        if ~isempty(scanpath_ws{1,1})
+            % Compute corresponding HMM
+            vbopt=initialize_HMM_computation(im);
+            [hmm_ws,~] = vbhmm_learn(scanpath_ws, K, vbopt);
+
+            % sort states from left to right
+            hmm_ws = sort_hmm_state(hmm_ws);
+
+            % add 'ghost states' if K < Kmax so all gaze descriptor vectors have the same dimension
+            hmm_ws=pad_with_ghost_states(hmm_ws,max(K),im);
+
+            %Extract gaze_descriptor vector from HMM parameters: priors, transition matrix coefficients, state centres and state covariances
+            gaze_descriptor_ws(isub,:) =extract_hmm_parameters(hmm_ws);
+
+            if isplotHMM
+                subplot(1,3,1)
+                plot_hmm_state(hmm_ws,scanpath_ws,im)
+                title WithSound
+            end
+        end
+        %% Auditory Condition 2: Without Orginal Soundtrack
+
+        % Extract current scanpath
+        scanpath_wos = extract_scanpath(example_EyeData_Coutrot,'without_os',isub,istim,K);
+        if ~isempty(scanpath_wos{1,1})
+            % Compute corresponding HMM
+            vbopt=initialize_HMM_computation(im);
+            vbopt.do_constrain_var=1;
+            [hmm_wos,~] = vbhmm_learn(scanpath_wos, K, vbopt);
+
+            % sort states from left to right
+            hmm_wos = sort_hmm_state(hmm_wos);
+
+            % add 'ghost states' if K < Kmax so all gaze descriptor vectors have the same dimension
+            hmm_wos=pad_with_ghost_states(hmm_wos,max(K),im);
+
+            %Extract gaze_descriptor vector from HMM parameters:
+            %priors, transition matrix coefficients, state centres and state covariances
+            gaze_descriptor_wos(isub,:) =extract_hmm_parameters(hmm_wos);
+
+            if isplotHMM
+                subplot(1,3,2)
+                plot_hmm_state(hmm_wos,scanpath_wos,im)
+                title WithoutSound
+                pause(0.1)
+            end
+        end
+    end
+
+    gaze_descriptor_ws(isnan(gaze_descriptor_ws(:,1)),:)=[];
+    gaze_descriptor_wos(isnan(gaze_descriptor_wos(:,1)),:)=[];
+
+    HMM_descriptor_Coutrot.(im_name_struct).with_os.gaze_descriptor=gaze_descriptor_ws;
+    HMM_descriptor_Coutrot.(im_name_struct).without_os.gaze_descriptor=gaze_descriptor_wos;
+
+end
+save('HMM_descriptor_Coutrot','HMM_descriptor_Coutrot')
+
+