HoughTransform/stats_ism_horses.m

function stats=stats_ism_horses
thresholds=[0.05:0.05:2];
method{1}='Hough';method{2}='PC/BC-DIM';%method{3}='baseline';method{4}='groundtruth';
numTest=228;
spread=10

%DEFINE CODEBOOK
load('ISM_codebook_horses.mat');
Locations=postprocess_codebook(Locations,patchClassList,spread);

%FOR EACH TEST IMAGE  PERFORM ISM TO LOCATE HORSE
disp('Processing Images')
for i=1:numTest*2
  fprintf(1,'.%i.',i);
  %load image and display
  if i<=numTest
    [I,groundtruth]=load_image_horse(100+i);%1st 100 images were used for training
  else
    I=load_image_101(100+i-numTest); groundtruth=I.*0;
    %I=load_image_car(100+i-numTest,2); groundtruth=I.*0;
  end
  I=padarray(I,[4,4],'replicate','both');
  groundtruth=padarray(groundtruth,[4,4],0,'both');
  figured(2),clf, imagesc(I), hold on, axis('equal','tight'); title(' '); 
  colormap('gray');
 
  %APPLY METHODS
  [param{i}{1},param{i}{2}]=ism_compare_methods(I,patches,patchClassList,Locations,similarityThres);
  trueBoundingBox{i}=groundtruth;

  if length(method)>=3, 
    param{i}{3}{1}(1:3,1)=[size(I')'./2;1]; %baseline - object in centre of image
  end
  if length(method)>=4, 
    param{i}{4}{1}(1:3,1)=[centroid(groundtruth)';1]; %groundtruth - object at correct position
  end
  %[~,la(i),lb(i)]=bounding_box(groundtruth); %measure bounding box size (need to change for-loop parameters to use training images)
end
%mean(la),mean(lb)

%FOR EACH METHOD AND THRESHOLD TEST ACCURACY WITH WHICH HORSES ARE LOCATED
for m=1:length(method)
  disp(' ');
  disp(method{m});
  
  disp('Testing Accuracy for threshold:')
  t=0;
  for thres=thresholds.*m %*m to make thresholds larger for DIM
    t=t+1;
    fprintf(1,'.%2.2f.',thres);
    stats{m}(t,:)=[0,0,0];
  
    %evaluate performance with each test image
    for i=1:numTest*2
      %find coordinates for which the votes exceed the threshold
      index=find(param{i}{m}{1}(3,:)>thres); 
      coords=param{i}{m}{1}(1:2,index);
      
      stats{m}(t,:)=horse_evaluation(stats{m}(t,:),coords,trueBoundingBox{i});
    end
    f1score{m}(t)=calc_f1score(stats{m}(t,:));
  end
end

%PLOT THE RESULTS
disp(' ');
max_error=0;
for m=1:length(method)
  max_error=max(max_error,max(sum(stats{m}(:,2:3),2)));
  disp(['f1score ',method{m}]);
  disp(f1score{m})
end
lineStyle{1}='b-s';lineStyle{2}='r-d';lineStyle{3}='g-o';lineStyle{4}='m-x';

figured(11),clf
for m=1:length(method)
  subplot(1,length(method),m),plot_errors(thresholds.*m,stats{m},max_error);title(method{m})
end

figured(12),clf
for m=1:length(method)
  plot_precision_recall(stats{m},lineStyle{m}), hold on
end
legend(method,'Location','southeast')

figured(13),clf
for m=1:length(method)
  plot_RFPPI(stats{m},lineStyle{m},numTest), hold on
end
legend(method,'Location','southeast')





function stats=horse_evaluation(stats,pos,groundtruth)
%assumes images contain exactly 1 or exactly zero objects, as overlap is calculated for whole image, not for multiple separate bounding boxes.
if sum(sum(groundtruth))>0, numObj=1; else, numObj=0; end


%make bounding box for groundtruth (padding image to ensure any predictions of box position will fit)
[a,b]=size(groundtruth);
pad=max(a,b);
groundtruth=padarray(groundtruth,[pad,pad],0,'both');
box_true=bounding_box(groundtruth);

%make bounding box around each predicted position - and calculate overlap with
%groundtruth bounding box. Note: average size of bounding box for the 100
%training images is 67x54 pixels
num=size(pos,2);
overlap=[];
for i=1:num
  y=pad+pos(2,i);
  x=pad+pos(1,i);
  box_predict=zeros(size(groundtruth));
  box_predict(floor(y-54/2):ceil(y+54/2),floor(x-67/2):ceil(x+67/2))=1;
  overlap(i)=sum(sum(box_predict.*box_true))./sum(sum(max(box_predict,box_true)));
end
overlap=max(overlap);

%update counts of true positives, false positives, and false negatives
if num==0
  TP=0;FP=0;FN=numObj;
elseif num==1
  if overlap>0.5
    TP=1;FP=0;FN=0;
  else
    TP=0;FP=1;FN=numObj;
  end    
else
  if overlap>0.5
    TP=1;FP=num-1;FN=0;
  else
    TP=0;FP=num;FN=numObj;
  end    
end
stats=stats+[TP,FP,FN];




function stats=horse_evaluation_dist(stats,pos,pos_true)
pos
num=size(pos,2);
dist=min(sqrt(sum(bsxfun(@minus,pos,pos_true).^2)));

thres=8;

if num==0
  TP=0;FP=0;FN=1;
elseif num==1
  if dist<thres
    TP=1;FP=0;FN=0;
  else
    TP=0;FP=1;FN=1;
  end    
else
  if dist<thres
    TP=1;FP=num-1;FN=0;
  else
    TP=0;FP=num;FN=1;
  end    
end
stats=stats+[TP,FP,FN];




function [box,la,lb]=bounding_box(I)
box=zeros(size(I));
[arange,brange]=find(I>0.5);
box(min(arange):max(arange),min(brange):max(brange))=1;
la=max(arange)-min(arange)+1;
lb=max(brange)-min(brange)+1;