lots of file updates

esvm_demo_train_synthetic.m

@@ -59,7 +59,7 @@
 
 val_params = params;
 val_params.detect_exemplar_nms_os_threshold = 0.5;
-val_params.gt_function = @get_pascal_anno_function;
+val_params.gt_function = @esvm_load_gt_function;
 val_set = cat(1, pos_set(:), val_neg_set(:));
 val_params.val_set = val_set;
 val_set_name = 'valset';

esvm_demo_train_voc_class_fast.m

@@ -39,9 +39,9 @@
 %results_directory = sprintf('/csail/vision-videolabelme/people/tomasz/esvm-%s/',cls);
 
 
-dataset_params = get_voc_dataset(dataset_directory,...
-                                 data_directory,...
-                                 results_directory);
+dataset_params = esvm_get_voc_dataset(dataset_directory,...
+                                      data_directory,...
+                                      results_directory);
 %dataset_params.display = 1;
 %dataset_params.dump_images = 1;
 
@@ -75,7 +75,7 @@
 e_stream_set = esvm_get_pascal_stream(stream_params, ...
                                       dataset_params);
 
-neg_set = get_pascal_set(dataset_params, ['train-' cls]);
+neg_set = esvm_get_pascal_set(dataset_params, ['train-' cls]);
 
 %Choose a models name to indicate the type of training run we are doing
 models_name = ...
@@ -94,19 +94,19 @@
 
 % val_params = params;
 % val_params.detect_exemplar_nms_os_threshold = 0.5;
-% val_params.gt_function = @get_pascal_anno_function;
+% val_params.gt_function = @esvm_load_gt_function;
 % val_params.CACHE_BETAS = 1;
 
 % val_set_name = ['trainval+' cls];
 
-% val_set = get_pascal_set(dataset_params, val_set_name);
+% val_set = esvm_get_pascal_set(dataset_params, val_set_name);
 % val_set = val_set(1:10);
 
 %% Define test-set
 test_params = params;
 test_params.detect_exemplar_nms_os_threshold = 0.5;
 test_set_name = ['trainval+' cls];
-test_set = get_pascal_set(dataset_params, test_set_name);
+test_set = esvm_get_pascal_set(dataset_params, test_set_name);
 test_set = test_set(1:100);
 
 %% Train the exemplars and get updated models name

esvm_pool_exemplar_dets.m

@@ -1,6 +1,6 @@
 function final = esvm_pool_exemplar_dets(grid, models, M, params)
-%% Perform detection post-processing and pool detection boxes
-%(which will then be ready to go into the PASCAL evaluation code)
+% Perform detection post-processing and pool detection boxes
+% (which will then be ready to go into the PASCAL evaluation code)
 % If there are overlap scores associated with boxes, then they are
 % also kept track of propertly, even after NMS.
 % 
@@ -65,21 +65,6 @@
 
 raw_boxes = bboxes;
 
-% perform within-exemplar NMS
-% NOTE: this is already done during detection time
-% if 0 
-%   fprintf(1,'applying exemplar nms\n');
-%   for i = 1:length(bboxes)
-%     if size(bboxes{i},1) > 0
-%       bboxes{i}(:,5) = 1:size(bboxes{i},1);
-%       bboxes{i} = nms_within_exemplars(bboxes{i},.5);
-%       if length(grid{i}.extras)>0 && isfield(grid{i}.extras,'os')
-%         maxos{i} = maxos{i}(bboxes{i}(:,5));
-%       end
-%     end
-%   end
-% end
-
 %Perform score rescaling
 %1. no scaling
 %2. platt's calibration (sigmoid scaling)

esvm_script_train_voc_class.m

@@ -34,9 +34,9 @@
 %data_directory = '/csail/vision-videolabelme/people/tomasz/VOCdevkit/';
 %results_directory = sprintf('/csail/vision-videolabelme/people/tomasz/esvm-%s/',cls);
 
-dataset_params = get_voc_dataset(dataset_directory, ...
-                                 data_directory, ...
-                                 results_directory);
+dataset_params = esvm_get_voc_dataset(dataset_directory, ...
+                                      data_directory, ...
+                                      results_directory);
 %dataset_params.display = 1;
 %dataset_params.dump_images = 1;
 
@@ -74,7 +74,7 @@
                                       dataset_params);
 
 
-neg_set = get_pascal_set(dataset_params, ['train-' cls]);
+neg_set = esvm_get_pascal_set(dataset_params, ['train-' cls]);
 
 %Choose a models name to indicate the type of training run we are doing
 models_name = ...
@@ -93,18 +93,18 @@
 
 val_params = params;
 val_params.detect_exemplar_nms_os_threshold = 0.5;
-val_params.gt_function = @get_pascal_anno_function;
+val_params.gt_function = @esvm_load_gt_function;
 val_params.CACHE_BETAS = 1;
 
 val_set_name = ['trainval'];
 
-val_set = get_pascal_set(dataset_params, val_set_name);
+val_set = esvm_get_pascal_set(dataset_params, val_set_name);
 
 %% Define test-set
 test_params = params;
 test_params.detect_exemplar_nms_os_threshold = 0.5;
 test_set_name = ['test'];
-test_set = get_pascal_set(dataset_params, test_set_name);
+test_set = esvm_get_pascal_set(dataset_params, test_set_name);
 
 %% Train the exemplars and get updated models name
 [models,models_name] = esvm_train_exemplars(initial_models, ...
@@ -122,11 +122,12 @@
 %% Apply calibration matrix to test-set results
 test_struct = esvm_pool_exemplar_dets(test_grid, models, M, test_params);
 
+%% Show top 20 detections as exemplar-inpainting results
 maxk = 20;
 allbbs = esvm_show_top_dets(test_struct, test_grid, test_set, models, ...
-                       params,  maxk, test_set_name);
+                            params,  maxk, test_set_name);
 
-[results] = evaluate_pascal_voc_grid(test_struct, test_grid,  ...
-                                     params, test_set_name, cls, ...
-                                     models_name);
+%% Perform the exemplar evaluation
+[results] = esvm_evaluate_pascal_voc(test_struct, test_grid, params, ...
+                                     test_set_name, cls, models_name);
 

esvm_train_exemplars.m

@@ -186,7 +186,7 @@
 
     %   figure(445);
     %   clf;
-    %   showI = get_sv_stack(m,5,5);
+    %   showI = esvm_show_det_stack(m,5,5);
     %   imagesc(showI);
     %   title('Exemplar and Top Dets');
     %   drawnow;

internal/esvm_estimate_M.m

@@ -56,8 +56,6 @@
 
 curcls = models{1}.cls;
 
-%curcls = find(ismember(params.dataset_params.classes,models{1}.cls));
-
 fprintf(1,' -Computing Box Features:');
 starter=tic;
 for i = 1:length(grid)
@@ -100,31 +98,10 @@
   end
 end
 
-
-
-if 0
-%% clip boxes to image
-fprintf(1,'clipping boxes\n');
-for i = 1:length(boxes)
-  boxes{i} = clip_to_image(boxes{i},grid{i}.imbb);
-end
-end
-
 lens = cellfun(@(x)size(x,1),boxes);
 boxes(lens==0) = [];
 maxos(lens==0) = [];
 
-%already nms-ed within exemplars (but not within LR flips)
-%%NOTE: should this be turned on?
-if 0
-  for i = 1:length(boxes)
-    boxes{i}(:,5) = 1:size(boxes{i},1);
-    boxes{i} = nms_within_exemplars(boxes{i},.5);
-    maxos{i} = maxos{i}(boxes{i}(:,5));
-    boxes{i}(:,5) = i;
-  end
-end
-
 K = length(models);
 N = sum(cellfun(@(x)size(x,2),maxos));
 
@@ -225,7 +202,6 @@
   %old way: works better!
   C(cur,exids(i)) = C(cur,exids(i)) + os(i)*(os(i) >= count_thresh) / ...
        length(cur);
-
 end
 
 for i = 1:K

internal/esvm_evaluate_pascal_voc.m

@@ -1,4 +1,4 @@
-function [results] = evaluate_pascal_voc_grid(test_struct, grid, ...
+function [results] = esvm_evaluate_pascal_voc(test_struct, grid, ...
                                               params, ...
                                               target_directory, cls, ...
                                               models_name)

internal/esvm_get_model_wiggles.m

@@ -1,5 +1,6 @@
-function model = populate_wiggles(I, model, NWIGGLES)
-%Get wiggles
+function model = esvm_get_model_wiggles(I, model, NWIGGLES)
+%Get wiggles of a model and populate model with them
+
 xxx = replica_hits(I, model.init_params.sbin, model.bb(1,:), ...
                    model.hg_size, NWIGGLES, model);
 
@@ -9,3 +10,124 @@
 model.w = model.w - mean(model.w(:));
 %model.b = -100;
 model.b = 0;
+
+function xxx = replica_hits(I, sbin, bb, hg_size, N_WIGGLES, savemodel)
+%% Given an image, an sbin, and the location of the detection window
+%% encoded as bb, create N_WIGGLES wiggles by perturbing image content
+%% inside the frame
+
+%% Tomasz Malisiewicz (tomasz@cmu.edu)
+
+if ~exist('N_WIGGLES','var')
+  N_WIGGLES = 100;
+end
+
+I = im2double(I);
+
+NEWPAD = 5; %ceil(max(hg_size(1),hg_size(2))/2);
+
+u = bb(9);
+v = bb(10);
+
+if bb(7) == 1
+  I = flip_image(I);
+end
+
+lilI = resize(I,bb(8));
+fprintf(1,'Warning: using esvm_features directly\n');
+f = esvm_features(lilI,sbin);
+f = padarray(f,[NEWPAD NEWPAD 0]);
+xxx_base = f(u-2+(1:hg_size(1)),v-2+(1:hg_size(2)),:);
+
+%%% NOW DO WIGGLES WITHOUT A SLIDE!
+xxx = zeros(prod(hg_size),N_WIGGLES);
+xxx(:,1) = xxx_base(:);
+for iii = 2:N_WIGGLES
+  %fprintf(1,'$');
+
+  %Add random noise
+  curI = lilI + .02*randn(size(lilI));
+
+  %Perform random gaussian smoothing
+  randsigma = .1+rand;
+  curI = (imfilter(curI,fspecial('gaussian',[9 9],randsigma)));
+
+  %Perform random shift
+  cx = floor(3*rand)-1;
+  cy = floor(3*rand)-1;
+  curI = circshift2(curI,[cx cy]);
+
+  % rrr = ceil(rand*4);
+  % if rrr == 1
+  %   TOP = floor(rand*size(curI,1)*.3);
+  %   curI(1:TOP,:,:) = 0;
+  % elseif rrr == 2
+  %   BOT = floor(rand*size(curI,1)*.3);
+  %   curI(end-BOT:end,:,:) = 0;
+  % elseif rrr == 3
+  %   LEFT = floor(rand*size(curI,2)*.3);
+  %   curI(:,1:LEFT,:) = 0;
+  % else
+  %   RIGHT = floor(rand*size(curI,2)*.3);
+  %   curI(:,end-RIGHT:end,:) = 0;
+  % end
+
+  f = esvm_features(curI,sbin);
+  f = padarray(f,[NEWPAD NEWPAD 0]);
+  x = f(u-2+(1:hg_size(1)),v-2+(1:hg_size(2)),:);
+  xxx(:,iii) = x(:);
+  if 0
+    figure(444)
+    clf
+    imagesc(max(0.0,min(1.0,curI)))
+    drawnow
+    pause
+  end
+end
+
+% return
+
+% clear sg;
+% cxs = 1:5:16;
+% cys = 1:5:16;
+% xxx = zeros(prod(size(w)),0);
+% for cxi = 1:length(cxs)
+%   for cyj = 1:length(cys)
+%     cx = cxs(cxi);
+%     cy = cys(cyj);
+
+%     curI = I(cx:end,cy:end,:);
+
+%     if rand>.5
+%       curI = rgb2gray(curI);
+%       curI = repmat(curI,[1 1 3]);
+%     end
+%     %curI = curI + .3*randn(size(curI));
+%     %curI = max(0.0,min(1.0,curI));
+%     randsigma = .1+rand*2;
+%     curI = (imfilter(curI,fspecial('gaussian',[9 9],randsigma)));
+%     curI = min(max(curI,0.0),1.0);
+
+%     [rs,t] = localizemeHOG(curI,...
+%                            sbin,...
+%                            {w},...
+%                            {b},...
+%                            -1.0,1,10,1);
+%     if 0
+%     mask = zeros(size(w,1),size(w,2));
+%     [coarse_boxes] = extract_bbs_from_rs(rs, curI, ...
+%                                           sbin, mask);
+%     bbb = coarse_boxes(1,:);
+%     (bbb(3)-bbb(1)+1) / (bbb(4)-bbb(2)+1)
+%     figure(44)
+%     clf
+%     imagesc(curI)
+%     plot_bbox(coarse_boxes(1,:))
+%     pause
+%     end
+%     fprintf(1,'\n');
+%     maxer=max(rs.score_grid{1});
+%     sg(cxi,cyj) = maxer;
+%     xxx(:,end+1) = rs.support_grid{1}{1};
+%   end
+% end

internal/esvm_get_pascal_set.m

@@ -1,4 +1,4 @@
-function bg = get_pascal_set(VOCopts, target_directory, class)
+function bg = esvm_get_pascal_set(VOCopts, target_directory, class)
 % Get a subset of PASCAL dataset (as defined in VOCopts) from the
 % 'train', 'trainval', or 'test' subsets ('both' refers to
 % 'trainval'+'test')
@@ -7,11 +7,18 @@
 % if class is "-car" chose only images not containing cars
 % returns:
 %   [bg]: a cell array of images, such that I=convert_to_I(bg{i});
+%
+% Copyright (C) 2011-12 by Tomasz Malisiewicz
+% All rights reserved.
+% 
+% This file is part of the Exemplar-SVM library and is made
+% available under the terms of the MIT license (see COPYING file).
+% Project homepage: https://github.com/quantombone/exemplarsvm
 
 if (nargin == 2) && strcmp(target_directory,'both')
   bg = cat(1, ...
-           get_pascal_set(VOCopts, 'trainval'),...
-           get_pascal_set(VOCopts, 'test'));
+           esvm_get_pascal_set(VOCopts, 'trainval'),...
+           esvm_get_pascal_set(VOCopts, 'test'));
   return;  
 end
 
@@ -27,7 +34,7 @@
 
   td = target_directory(1:has_marker(1)-1);
   cl = target_directory(has_marker(1):end);
-  bg = get_pascal_set(VOCopts, td, cl);
+  bg = esvm_get_pascal_set(VOCopts, td, cl);
   return;
 end
 
@@ -65,7 +72,7 @@
   end
   neg_set = neg_set(gt==1);
 else
-  error(sprintf(['Invalid class %s passed to get_pascal_set: must' ...
+  error(sprintf(['Invalid class %s passed to esvm_get_pascal_set: must' ...
                  ' start with plus or minus sign'], class));
 end