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。

function do_test_efficient(config_file)
  
 %% Function that takes a hand-trained parts and structure model and the
%% interest points produced by running the part templates over the images,
%% and finding the best configuration of them in each image.

%% Based on method in the paper:
%% Felzenszwalb, P. and Huttenlocher, D. "Pictorial Structures for Object
%% Recognition", Intl. Journal of Computer Vision, 61(1), pp. 55-79, January 2005.    
    
%% Does this for all images and then plots training and test separately.  
  
%% Before running this, you must have run:
%%    do_random_indices - to generate random_indices.mat file
%%    do_preprocessing - to get the images that the operator will run on  
%%    do_manual_train_parts_structure - to get the model and part
%%                                      templates.
%%    do_parts_filtering - to get interest points from part templates.
  
%% R.Fergus ([e-mail]) 03/10/05.  

%%% debug switch
DEBUG = 0;

%%% base figure number
BASE_FIG_NUM = 500;

%%% standard color ordering
cols = {'r.' 'g.' 'b.' 'c.' 'm.' 'y.' 'k.'};
cols2 = {'rx' 'gx' 'bx' 'cx' 'mx' 'yx' 'kx'};
cols3 = {'ro' 'go' 'bo' 'co' 'mo' 'yo' 'ko'};
cols4 = {'r' 'g' 'b' 'c' 'm' 'y' 'k'};

%% Evaluate global configuration file
eval(config_file);

%% Get list of file name of input images of positive class only
img_file_names = genFileNames({Global.Image_Dir_Name},[1:Categories.Total_Frames],RUN_DIR,Global.Image_File_Name,Global.Image_Extension,Global.Num_Zeros);

%% Get list of output file names
ip_file_names =  genFileNames({Global.Interest_Dir_Name},[1:Categories.Total_Frames],RUN_DIR,Global.Interest_File_Name,'.mat',Global.Num_Zeros);

%% load up most recent model and take model from it.... 

%%% just take newest model in subdir.
ind = length(dir([RUN_DIR,'/',Global.Model_Dir_Name,'/', Global.Model_File_Name,'*.mat']));    
    
%%% construct model file name
model_fname = [RUN_DIR,'/',Global.Model_Dir_Name,'/',Global.Model_File_Name,prefZeros(ind,Global.Num_Zeros),'.mat']

%%% load up model
load(model_fname);

%%% Get total number of images
nImages = length(ip_file_names);

%%% get # parts
nParts = Learn.Num_Parts;

if (~exist('bounding_box_efficient'))

%%% precompute shape density variance
for b=2:Learn.Num_Parts
  determinant_term(b) = -log(shape_var_x(b) * shape_var_y(b) * 2 * pi); 
end

%%% round means to integer values
shape_mean_x = round(shape_mean_x);
shape_mean_y = round(shape_mean_y);

%%% get absolute values of means
abs_mean_x = abs(shape_mean_x);
abs_mean_y = abs(shape_mean_y);


tic;

for i = 1:nImages  
  
  %%% display progress
  if (mod(i,10)==0)
    fprintf('.%d',i);
  end
  
  %%% load up interest files 
  load(ip_file_names{i});
    
  %%% get size of response_image
  [imy,imx,tmp] = size(response_image);
  
  %%% setup variables
  overall_resp = zeros(imy,imx);
  part_resp = zeros(imy,imx,nParts-1);
  locations = cell(1,nParts-1);

  for p=2:Learn.Num_Parts
    
     %% Pad response_image with v.large costs (not Inf since it screws up
     %% dist_transform_1d function). This is designed to ensure we don't
     %% run out of image when we translate it 
	    
     %% Also: (a) scale Appearance costs, giving weighting between shape and
     %% appearance, and (b) negate everything since for distance transform, low is
     %% good but probability map has good being high values_efficient
     padded_response_image = padarray(-Recog.Shape_Appearance_Weighting*response_image(:,:,p),[abs_mean_y(p) abs_mean_x(p)],1e200,'both');                      
	   
     %%% translate response image by mean of that part relative to the
     %%% landmark
     translated_response = padded_response_image([abs_mean_y(p)+shape_mean_y(p)+1:imy+(abs_mean_y(p)+shape_mean_y(p))],[abs_mean_x(p)+shape_mean_x(p)+1:imx+(abs_mean_x(p)+shape_mean_x(p))]);
     
     %%% Now do distance transform, using variance as multiplication factor.
     [part_resp(:,:,p-1),locations{p-1}] = dist_transform(translated_response,shape_var_x(p),shape_var_y(p));
	 
 
  end

  %% Now add in landmark response to sum of all other parts...
  overall_resp = (-Recog.Shape_Appearance_Weighting * response_image(:,:,1)) + sum(part_resp,3);

  %% Find minimum overall 
  [min_score,landmark_pos] = min(overall_resp(:));
  
  %% Get landmark x,y location
  [landmark_y, landmark_x] = ind2sub([imy imx],landmark_pos);

  %% Get locations of other parts...
  for p = 2:nParts
    part_x(p-1) = locations{p-1}(landmark_y,landmark_x,1);
    part_y(p-1) = locations{p-1}(landmark_y,landmark_x,2);
  end
	 
  %%% Overall best locations....
  best_x = [landmark_x,part_x+shape_mean_x(2:end)'];
  best_y = [landmark_y,part_y+shape_mean_y(2:end)'];
  
  %%% plot stufff
  if DEBUG
    

    figure(BASE_FIG_NUM); clf;
    im = imread(img_file_names{i});
    imagesc(im); hold on;
%   
    for p=1:nParts
      plot(best_x(p),best_y(p),cols{p},'Markersize',20,'Linewidth',5);
    end
    
    figure(BASE_FIG_NUM+1); clf;
    %%% landmark response
    subplot(2,nParts,nParts+1);
    imagesc(-overall_resp); caxis([-1 1]*Recog.Shape_Appearance_Weighting);
    hold on; title('Overall probab.');
    plot(landmark_x,landmark_y,'wx','Markersize',20,'Linewidth',5);
  
    %%% get best overall reponse
    q = -response_image(:,:,1);
    [tmp,pos] = min(q(:));
    [yy,xx] = ind2sub([imy imx],pos);
   % plot(xx,yy,'ko','Markersize',20,'Linewidth',5);
    colorbar;
    
    subplot(2,nParts,1);
    imagesc(response_image(:,:,1)); caxis([-1 1]); colorbar;
    title(['Appearance probab., part 1']); hold on;
    plot(xx,yy,'ko','Markersize',20,'Linewidth',5);
    
    %%% non-landmark parts
    for p=2:nParts
    %%% get best overall reponse

      q = -response_image(:,:,p);
      [tmp,pos] = min(q(:));
      [yy,xx] = ind2sub([imy imx],pos);
   
      subplot(2,nParts,nParts+p)
      imagesc(-part_resp(:,:,p-1)); hold on; caxis([-1 1]*Recog.Shape_Appearance_Weighting); colorbar;      
      plot(part_x,part_y,'wx','Markersize',20,'Linewidth',5);
      plot(xx-shape_mean_x(p),yy-shape_mean_y(p),'ko','Markersize',20,'Linewidth',5);
      title(['Distance transformed part ',num2str(p)]);
      subplot(2,nParts,p);
      imagesc(response_image(:,:,p)); hold on; caxis([-1 1]); colorbar;
      plot(xx,yy,'ko','Markersize',20,'Linewidth',5);
      title(['Appearance probab., part ',num2str(p)]);
    end

    %%% plot best 

    pause
  end

  
  %%% Collect statistics for plotting and ROC and RPC curves
  best_locations_x(:,i) = best_x';
  best_locations_y(:,i) = best_y';
  
  %% now find min and max x and y
  x_min = min(best_x);
  y_min = min(best_y);
  x_max = max(best_x);
  y_max = max(best_y);
  %% now compute bounding rectangle
  bounding_box_efficient{i} = round([x_min,y_min,x_max-x_min,y_max-y_min])';
  %% get score..
  best_overall_score_efficient(i) = -min_score; %% negate to turn cost back into probability... 
  
end

total_time=toc;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Compute some performance metrics
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%%% get labels for test frames
labels_efficient = zeros(1,Categories.Total_Frames);
for a=1:Categories.Number
  labels_efficient(Categories.Test_Frames{a}) = [Categories.Labels(a)*ones(1,length(Categories.Test_Frames{a}))];
end