open source pkg v1

pkg/OpenFace/matlab_runners/Gaze Experiments/extract_columbia_gaze_test.m

@@ -0,0 +1,97 @@
+clear
+     
+curr_dir = cd('.');
+
+database_root = ['D:\Datasets\Columbia Gaze Data Set\'];
+p_dirs = dir([database_root, '0*']);
+
+output = './columbia_out/';
+
+%% Perform actual gaze predictions
+command = sprintf('"../../Release/FaceTrackingImg.exe" -fx 19720 -fy 19720 -gaze ');
+
+parfor p=1:numel(p_dirs)
+
+    input_loc = ['-fdir "', [database_root, p_dirs(p).name], '" '];
+    out_img_loc = ['-oidir "', [output, p_dirs(p).name], '" '];
+    out_p_loc = ['-opdir "', [output, p_dirs(p).name], '" '];
+    command_c = cat(2, command, input_loc, out_img_loc, out_p_loc);
+
+    command_c = cat(2, command_c, ' -clmwild -multi_view 1');
+    dos(command_c);
+    
+end
+
+%% Perform the evaluation (this needs to be changed)
+errors_l = [];
+errors_r = [];
+
+all_gaze_pred = [];
+all_gaze_gt = [];
+angle_errs = [];
+
+angle_errs_naive = [];
+
+frontal_faces = [];
+
+% It's approximate as the actual eye gaze is approximate
+for p=1:numel(p_dirs)
+   
+    out_files = dir([output, p_dirs(p).name, '/*.pose']);
+    
+    for i=1:numel(out_files)
+        out_file = [output, p_dirs(p).name, '/', out_files(i).name];
+        A = dlmread(out_file, ' ', 'A79..F79');       
+        g_0 = A(1:3);
+        g_1 = A(4:6);
+                
+        g_0 = (g_0 + g_1) / 2;
+        g_0 = g_0 ./ norm(g_0);
+        g_1 = g_0;
+        
+        
+        all_gaze_pred = cat(1, all_gaze_pred, g_0);
+        all_gaze_pred = cat(1, all_gaze_pred, g_1);        
+        
+        tokens = strsplit(out_files(i).name, '_');
+        
+        gaze_v_gt = str2double(tokens{4}(1:end-1)) * pi/180;
+        gaze_h_gt = str2double(tokens{5}(1:end-1)) * pi/180;
+        
+        person_h_gt = str2double(tokens{3}(1:end-1)) * pi/180;
+        
+        if(person_h_gt == 0)
+            frontal_faces = cat(1, frontal_faces, [1;1]);
+        else
+            frontal_faces = cat(1, frontal_faces, [0;0]);
+        end
+        
+        p_target_x = tan(gaze_h_gt) * 2500;
+        cam_dist = sqrt(2500^2 + p_target_x^2);                
+        
+        gaze_target_x = tan(gaze_h_gt) * cam_dist;
+        gaze_target_y = -tan(gaze_v_gt) * cam_dist;
+        
+        gaze_gt_1 = [gaze_target_x, gaze_target_y, 0] - [0,0,cam_dist];
+        gaze_gt_1 = gaze_gt_1 ./ norm(gaze_gt_1);
+        
+        gaze_gt_2 = [gaze_target_x, gaze_target_y, 0] - [0,0,cam_dist];
+        gaze_gt_2 = gaze_gt_2 ./ norm(gaze_gt_2);
+        
+        all_gaze_gt = cat(1, all_gaze_gt, gaze_gt_1);
+        all_gaze_gt = cat(1, all_gaze_gt, gaze_gt_2);
+        
+        % Gaze gt needs to be rotated based on person location
+        
+        angle_err_1 = acos(gaze_gt_1 * g_0') * 180/pi;
+        angle_err_2 = acos(gaze_gt_2 * g_1') * 180/pi;
+        angle_errs = cat(1, angle_errs, angle_err_1);
+        angle_errs = cat(1, angle_errs, angle_err_2);
+        
+        angle_errs_n_1 = acos(gaze_gt_1 * [0;0;-1]) * 180/pi;
+        angle_errs_n_2 = acos(gaze_gt_2 * [0;0;-1]) * 180/pi;
+        angle_errs_naive = cat(1, angle_errs_naive, angle_errs_n_1);
+        angle_errs_naive = cat(1, angle_errs_naive, angle_errs_n_2);
+    end
+    
+end

pkg/OpenFace/matlab_runners/Gaze Experiments/extract_mpii_gaze_test.m

@@ -0,0 +1,114 @@
+clear
+     
+curr_dir = cd('.');
+
+% Replace this with your downloaded 300-W train data
+if(exist([getenv('USERPROFILE') '/Dropbox/AAM/eye_clm/mpii_data/'], 'file'))
+    database_root = [getenv('USERPROFILE') '/Dropbox/AAM/eye_clm/mpii_data/'];    
+elseif(exist('D:\Dropbox/Dropbox/AAM/eye_clm/mpii_data/', 'file'))
+    database_root = 'D:\Dropbox/Dropbox/AAM/eye_clm/mpii_data/';    
+elseif(exist('F:\Dropbox/AAM/eye_clm/mpii_data/', 'file'))
+    database_root = 'F:\Dropbox/AAM/eye_clm/mpii_data/';    
+elseif(exist('/multicomp/datasets/mpii_gaze/mpii_data/', 'file'))
+    database_root = '/multicomp/datasets/mpii_gaze/mpii_data/';    
+elseif(exist('/media/tadas/5E08AE0D08ADE3ED/Dropbox/AAM/eye_clm/mpii_data/', 'file'))
+    database_root = '/media/tadas/5E08AE0D08ADE3ED/Dropbox/AAM/eye_clm/mpii_data/';
+else
+    fprintf('MPII gaze dataset not found\n');
+end
+
+output = './mpii_out/';
+
+%% Perform actual gaze predictions
+if(isunix)
+    executable = '"../../build/bin/FaceLandmarkImg"';
+else
+    executable = '"../../x64/Release/FaceLandmarkImg.exe"';
+end
+
+command = sprintf('%s -fx 1028 -fy 1028  ', executable);
+p_dirs = dir([database_root, 'p*']);
+
+parfor p=1:numel(p_dirs)
+    tic
+
+    input_loc = ['-gaze -fdir "', [database_root, p_dirs(p).name], '" '];
+    out_img_loc = ['-out_dir "', [output, p_dirs(p).name], '" '];
+    command_c = cat(2, command, input_loc, out_img_loc);
+
+    if(isunix)
+        unix(command_c, '-echo');
+    else
+        dos(command_c);
+    end
+
+end
+%%
+
+% Extract the results
+predictions_l = zeros(750, 3);
+predictions_r = zeros(750, 3);
+gt_l = zeros(750, 3);
+gt_r = zeros(750, 3);
+
+angle_err_l = zeros(750,1);
+angle_err_r = zeros(750,1);
+
+p_dirs = dir([database_root, 'p*']);
+curr = 1;
+for p=1:numel(p_dirs)
+    load([database_root, p_dirs(p).name, '/Data.mat']);
+
+    for i=1:size(filenames, 1)
+
+        fname = sprintf('%s/%s/%d_%d_%d_%d_%d_%d_%d.csv', output, p_dirs(p).name,...
+             filenames(i,1), filenames(i,2), filenames(i,3), filenames(i,4),...
+             filenames(i,5), filenames(i,6), filenames(i,7));
+        
+        if(p==1 && i==1)
+            % First read in the column names
+            tab = readtable(fname);
+            column_names = tab.Properties.VariableNames;
+
+            gaze_0_ids = cellfun(@(x) ~isempty(x) && x==1, strfind(column_names, 'gaze_0_'));
+            gaze_1_ids = cellfun(@(x) ~isempty(x) && x==1, strfind(column_names, 'gaze_1_'));
+        end
+        
+        if(exist(fname, 'file'))
+            all_params  = dlmread(fname, ',', 1, 0);
+        else
+            all_params = [];
+        end
+        
+        % If there was a face detected
+        if(size(all_params,1)>0)
+            predictions_r(curr,:) = all_params(1,gaze_0_ids);
+            predictions_l(curr,:) = all_params(1,gaze_1_ids);
+        else
+            predictions_r(curr,:) = [0,0,-1];
+            predictions_l(curr,:) = [0,0,-1];            
+        end
+
+        head_rot = headpose(i,1:3);
+ 
+        gt_r(curr,:) = data.right.gaze(i,:)';
+        gt_r(curr,:) = gt_r(curr,:) / norm(gt_r(curr,:));
+        gt_l(curr,:) = data.left.gaze(i,:)';
+        gt_l(curr,:) = gt_l(curr,:) / norm(gt_l(curr,:));
+
+        angle_err_l(curr) = acos(predictions_l(curr,:) * gt_l(curr,:)') * 180/pi;
+        angle_err_r(curr) = acos(predictions_r(curr,:) * gt_r(curr,:)') * 180/pi;
+
+        curr = curr + 1;
+    end
+
+end
+all_errors = cat(1, angle_err_l, angle_err_r);
+mean_error = mean(all_errors);
+median_error = median(all_errors);
+save('mpii_1500_errs.mat', 'all_errors', 'mean_error', 'median_error');
+
+f = fopen('mpii_1500_errs.txt', 'w');
+fprintf(f, 'Mean error, median error\n');
+fprintf(f, '%.3f, %.3f\n', mean_error, median_error);
+fclose(f);

pkg/OpenFace/matlab_runners/Gaze Experiments/mpii_1500_errs.txt

@@ -0,0 +1,2 @@
+Mean error, median error
+9.072, 8.454