do_test_efficient.m

一个学习自然场景类别的贝叶斯模型、基于“词袋”模型的目标分类。来源于Feifei Li的论文。是近年来的目标识别模型热点之一。

%%% now get scores of test images
values_efficient = -Inf * ones(1,Categories.Total_Frames);
for a=1:Categories.Number
  values_efficient(Categories.Test_Frames{a}) = best_overall_score_efficient(Categories.Test_Frames{a});
end

%%% since all training images are given -Inf score, weed out to avoid
%%% skewing ROC curve
good_ind = find(~isinf(values_efficient));
values_efficient2 = values_efficient(good_ind); labels_efficient2 = labels_efficient(good_ind);

%%% Now compute object present/absent peformance using ROC curve
[roc_curve_efficient,roc_op_efficient,roc_area_efficient,roc_threshold_efficient] = roc([values_efficient2;labels_efficient2]');

%%% Now do localization performance. This will be measured only on +ve
%categories, since all proposed detections on -ve data are definately
%false alarms. Note that the ground_truth_locations_{category_name} files
%produced by do_preprocessing.m must exist in RUN_DIR.
%% load up 

%%% first rescale proposed bounding box...
for a=1:length(bounding_box_efficient)

  %%% Enlarge proposed bounding box by Recog.Manual_Bounding_Box_Efficient_Scaling
  %%% first get centroid
  centroid_x = bounding_box_efficient{a}(1,:) + bounding_box_efficient{a}(3,:)/2;
  centroid_y = bounding_box_efficient{a}(2,:) + bounding_box_efficient{a}(4,:)/2;   
  
  %%% repoistion new top left corner
  bounding_box_efficient{a}(1,:) = centroid_x - bounding_box_efficient{a}(3,:)/2*Recog.Manual_Bounding_Box_Scaling;
  bounding_box_efficient{a}(2,:) = centroid_y - bounding_box_efficient{a}(4,:)/2*Recog.Manual_Bounding_Box_Scaling;
  
  %%% adjust width and height
  bounding_box_efficient{a}(3,:) = bounding_box_efficient{a}(3,:) * Recog.Manual_Bounding_Box_Scaling;
  bounding_box_efficient{a}(4,:) = bounding_box_efficient{a}(4,:) * Recog.Manual_Bounding_Box_Scaling;    
    
end

localization_labels_efficient = zeros(1,Categories.Total_Frames);
gt_boxes = cell(1,Categories.Total_Frames);

total_number_instances = 0;
for a=1:Categories.Number
  if (Categories.Labels(a)==1)
    fname = [RUN_DIR,'/',Global.Ground_Truth_Name,'_',Categories.Name{a},'.mat'];
    if (exist(fname))
      load(fname);
      
      %%% get relavent boxes
      gt_boxes(Categories.Test_Frames{a}) = gt_bounding_boxes(:,Categories.Test_Frames{a});
       
      %%% record total number of instances
      total_number_instances = total_number_instances + sum(cellfun('size',gt_boxes,2));
    else
      error([fname,' does not exist']);
    end
    
    %% find correct localizations
    localization_labels_efficient(Categories.Test_Frames{a}) = test_localization(gt_boxes(Categories.Test_Frames{a}),bounding_box_efficient(Categories.Test_Frames{a}),Recog);

  else
    %% all putative detection on negatively labelled images are
    %% automatically false alarms....
     
    localization_labels_efficient(Categories.Test_Frames{a}) = zeros(1,length(Categories.Test_Frames{a}));
    
  end 
end

%% Now compute recall_prescision curve
[rpc_curve_efficient,rpc_ap_efficient,rpc_area_efficient,rpc_threshold_efficient] = recall_precision_curve([values_efficient;localization_labels_efficient]',total_number_instances);

%%% Now save to model
save(model_fname,'best_locations_x','best_locations_y','best_overall_score_efficient','bounding_box_efficient','gt_boxes','roc_curve_efficient','roc_area_efficient','roc_op_efficient','roc_threshold_efficient','rpc_curve_efficient','rpc_ap_efficient','rpc_area_efficient','rpc_threshold_efficient','values_efficient','localization_labels_efficient','labels_efficient','-append');


fprintf('\nFinished running parts and structure model, using efficient methods, over images\n');
fprintf('Total number of images: %d, mean time per image: %f secs\n',Categories.Total_Frames,total_time/Categories.Total_Frames);




end %% end of computing results section

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Plotting section
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%% We will use figures from FIGURE_BASE to FIGURE_BASE + 4;
%% clear them ready for plotting action...
for a=BASE_FIG_NUM:BASE_FIG_NUM+2
    figure(a); %clf;
end

figure(BASE_FIG_NUM);
q=get(BASE_FIG_NUM,'UserData');
if isempty(q)
  plot(roc_curve_efficient(:,1),roc_curve_efficient(:,2),cols4{1}); hold on;
  set(BASE_FIG_NUM,'UserData',1);
else
  q=q+1;
  plot(roc_curve_efficient(:,1),roc_curve_efficient(:,2),cols4{rem(q-1,7)+1}); hold on;
  set(BASE_FIG_NUM,'UserData',q);
end
axis([0 1 0 1]); axis square; grid on;
xlabel('P_{fa}'); ylabel('P_d'); title(['ROC Curve, Area: ',num2str(roc_area_efficient),' OpP: ',num2str(roc_op_efficient)]);

figure(BASE_FIG_NUM+1);
q=get(BASE_FIG_NUM+1,'UserData');
if isempty(q)
  plot(rpc_curve_efficient(:,1),rpc_curve_efficient(:,2),cols4{1}); hold on;  
  set(BASE_FIG_NUM+1,'UserData',1);
else
  q=q+1;
  plot(rpc_curve_efficient(:,1),rpc_curve_efficient(:,2),cols4{rem(q-1,7)+1}); hold on;
  set(BASE_FIG_NUM+1,'UserData',q);
end
axis([0 1 0 1]); axis square; grid on;
xlabel('Recall'); ylabel('Precision'); title('RPC Curves');

%% first decide on plotting order
if strcmp(Plot.Example_Mode,'ordered')
    %%% just go in orginial order of images
    plot_order = sort(Categories.All_Test_Frames);
elseif strcmp(Plot.Example_Mode,'alternate')
    %%% using random order but alternating between images of different
    %%% classes...
    ind = ones(Categories.Number,max(cellfun('length',Categories.Test_Frames)));
    tmp = length(Categories.Test_Frames{1});
    ind(1,1:tmp)=[1:tmp];
    for a=2:Categories.Number
        tmp = length(Categories.Test_Frames{a});
        offset=sum(cellfun('length',Categories.Test_Frames(1:a-1)));
        ind(a,1:tmp) = [1:tmp]+offset;
   end
   plot_order =  Categories.All_Test_Frames(ind(:)); 
elseif strcmp(Plot.Example_Mode,'random')
    %%% using order given in random_indices.mat
    plot_order = Categories.All_Test_Frames;
elseif strcmp(Plot.Example_Mode,'best')
    %%% plot ordered by score 
    [tmp2,ind] =(sort(-best_overall_score_efficient(Categories.All_Test_Frames)));
    plot_order =  Categories.All_Test_Frames(ind);
elseif strcmp(Plot.Example_Mode,'worst')
    %%% plot ordered by score on worst topic
    [tmp2,ind] =(sort(best_overall_score_efficient(Categories.All_Test_Frames)));
    plot_order =  Categories.All_Test_Frames(ind);    
elseif strcmp(Plot.Example_Mode,'borderline')
    %%% ordering by how close they are to the topic_thresholds...
    [tmp2,ind] = sort(abs(best_overall_score_efficient(Categories.All_Test_Frames)-roc_threshold_efficient));
    plot_order =  Categories.All_Test_Frames(ind);  
else
    error('Unknown type of Plot.Example_Mode');
end 

%% now setup figure and run loop plotting images
figure(BASE_FIG_NUM+2);
nImage_Per_Figure = prod(Plot.Number_Per_Figure);

for a=1:nImage_Per_Figure:length(Categories.All_Test_Frames)
    
    clf; %% clear figure
    
    for b=1:nImage_Per_Figure
        
        %%% actual index
        index = plot_order(a+b-1);
        
        %%% get correct subplot
        subplot(Plot.Number_Per_Figure(1),Plot.Number_Per_Figure(2),b);
        
        %%% load image
        im=imread(img_file_names{index});
        
        %%% show image
        imagesc(im); hold on;
        
        %%% if grayscale, then adjust colormap
        if (size(im,3)==1)
            colormap(gray);
        end 
        
        %%% load up interest_point file
        load(ip_file_names{index});
        
        %%% now mark in best hypothesis
        for b=1:Learn.Num_Parts
          plot(best_locations_x(b,index),best_locations_y(b,index),cols2{b},'Markersize',20,'Linewidth',8);
        end

        
        if (~isempty(gt_boxes{index}))
          %%% show ground_truth bounding box
          rectangle('Position',gt_boxes{index},'EdgeColor','b','Linewidth',2);
          
          if (localization_labels_efficient(index)==1)
            %% correct localization
            rectangle('Position',bounding_box_efficient{index},'EdgeColor','g','Linewidth',2);
          else
            %% incorrect
            rectangle('Position',bounding_box_efficient{index},'EdgeColor','r','Linewidth',2);
          end
        end
        
        h=title(['Image: ',num2str(index),' Best match score: ',num2str(best_overall_score_efficient(index))]);
        if (labels_efficient(index)==1)
          set(h,'Color','g');
        else
          set(h,'Color','r');          
        end
        
    end
    
    pause
    
end