open source pkg v1

pkg/OpenFace/matlab_runners/Head Pose Experiments/calcBUerror.m

@@ -0,0 +1,82 @@
+function [meanError, all_rot_preds, all_rot_gts, all_errors, rels_all, seq_ids] = calcBUerror(resDir, gtDir)
+
+seqNames = {'jam1','jam2','jam3','jam4','jam5','jam6','jam7','jam8','jam9', ...
+    'jim1','jim2','jim3','jim4','jim5','jim6','jim7','jim8','jim9', ...
+    'llm1','llm2','llm3','llm4','llm5','llm6','llm7','llm8','llm9', ...
+    'ssm1','ssm2','ssm3','ssm4','ssm5','ssm6','ssm7','ssm8','ssm9', ...
+    'vam1','vam2','vam3','vam4','vam5','vam6','vam7','vam8','vam9'};
+
+rotMeanErr = zeros(numel(seqNames),3);
+rotRMS = zeros(numel(seqNames),3);
+rot = cell(1,numel(seqNames));
+rotg = cell(1,numel(seqNames));
+rels_all = [];
+
+seq_ids = {};
+
+for i = 1:numel(seqNames)
+    fname = [resDir seqNames{i} '.csv'];
+    if(i == 1)
+        % First read in the column names
+        tab = readtable(fname);
+        column_names = tab.Properties.VariableNames;
+
+        confidence_id = cellfun(@(x) ~isempty(x) && x==1, strfind(column_names, 'confidence'));
+        rot_ids = cellfun(@(x) ~isempty(x) && x==1, strfind(column_names, 'pose_R'));
+    end
+
+    all_params  = dlmread(fname, ',', 1, 0);
+    
+    rot{i} = all_params(:, rot_ids);    
+    rels = all_params(:, confidence_id);
+    
+    posesGround =  load ([gtDir seqNames{i} '.dat']);
+
+    % the reliabilities of head pose
+    rels_all = cat(1, rels_all, rels);
+    
+    % Flip because of different conventions
+    rot{i}(:,2) = -rot{i}(:,2);
+    rot{i}(:,3) = -rot{i}(:,3);
+    
+    % Convert to radians
+    rotg{i} = posesGround(2:end,[7 6 5]) * (pi/180);
+        
+    % Correct the first frame so it corresponds to (0,0,0), as slightly
+    % different pose might be assumed frontal and this corrects for
+    % that
+    
+    % Work out the correction matrix for ground truth
+    rot_corr_gt = Euler2Rot(rotg{i}(1,:));
+    for r_e = 1:size(rotg{i},1)
+        rot_curr_gt = Euler2Rot(rotg{i}(r_e,:));
+        rot_new_gt = rot_corr_gt' * rot_curr_gt;
+        rotg{i}(r_e,:) = Rot2Euler(rot_new_gt);
+    end
+    
+    % Work out the correction matrix for estimates
+    rot_corr_est = Euler2Rot(rot{i}(1,:));
+    for r_e = 1:size(rot{i},1)
+        rot_curr_est = Euler2Rot(rot{i}(r_e,:));
+        rot_new_est = rot_corr_est' * rot_curr_est;
+        rot{i}(r_e,:) = Rot2Euler(rot_new_est);
+    end
+    
+    % Convert to degrees
+    rotg{i} = rotg{i} * 180 / pi;
+    rot{i} = rot{i} * 180 / pi;  
+    
+    rotMeanErr(i,:) = mean(abs((rot{i}(:,:)-rotg{i}(:,:))));
+    rotRMS(i,:) = sqrt(mean(((rot{i}(:,:)-rotg{i}(:,:))).^2)); 
+    seq_ids = cat(1, seq_ids, repmat(seqNames(i), size(rot{i},1), 1));
+end
+allRot = cell2mat(rot');
+allRotg = cell2mat(rotg');
+
+meanError = mean(abs((allRot(:,:)-allRotg(:,:))));
+all_errors = abs(allRot-allRotg);
+rmsError = sqrt(mean(((allRot(:,:)-allRotg(:,:))).^2)); 
+errorVariance = var(abs((allRot(:,:)-allRotg(:,:))));      
+
+all_rot_preds = allRot;
+all_rot_gts = allRotg;