function [options , Xfa] = train_stage_cascade(Xtrain , ytrain , Xtest , ytest , options)
%
%
% Train model for a stage of the cascade
%
% Usage
% ------
%
% [options , Xfa] = train_stage_cascade(Xtrain , ytrain , Xtest , ytest , options)
%
%
% Inputs
% -------
%
%
% Xtrain Train Data matrix ( Ny x Nx x (options.Npostrain+options.Nnegtrain))
% with postives and negatives examples for the current
% stage to train
%
% ytrain Train labels (1 x (options.Npostrain+options.Nnegtrain))
%
% Xtest Trest Data matrix ( Ny x Nx x (options.Npostest+options.Nnegtest))
% with postives and negatives examples for testing current
% trained stage
%
% ytest Test labels (1 x (options.Npostest+options.Nnegtest))
%
% options Input options struture (see train_cascade function)
%
%
% Outputs
% -------
%
% options Updated Options struture (see train_cascade function)
%
% Xfa False alarms (non-faces above threshold) (Ny x Nx x Nfa)
%
%
% Author : Sbastien PARIS : sebastien.paris@lsis.org
% ------- Date : 01/27/2009
%
% Ref : [1] M-T. Pham and all, "Detection with multi-exit asymetric boosting", CVPR'08
% ------ [2] P.A Viola and M. Jones, "Robust real-time face detection",
% International Journal on Computer Vision, 2004
%
%
Ntrain = size(Xtrain , 3);
Ntest = size(Xtest , 3);
if(options.cascade_type == 1)
if(options.typefeat == 0)
fxtrain = eval_haar(Xtrain , options);
if(options.algoboost < 3)
IItrain = image_integral_standard(Xtrain);
else
Htrain = haar(Xtrain , options);
end
elseif(options.typefeat == 1)
fxtrain = eval_mblbp(Xtrain , options);
Htrain = mblbp(Xtrain , options);
end
ctelambda = -0.5*log(options.lambda);
indptrain =(ytrain == 1);
indntrain = ~indptrain;
wtrain = zeros(1 , Ntrain);
wtrain(indptrain) = exp(-(fxtrain(indptrain) + ctelambda))/options.Npostrain;
wtrain(indntrain) = exp(fxtrain(indntrain) + ctelambda)/options.Nnegtrain;
elseif(options.cascade_type == 0)
if(options.typefeat == 0)
IItrain = image_integral_standard(Xtrain);
elseif(options.typefeat == 1)
Htrain = mblbp(Xtrain , options);
end
wtrain = ones(1,Ntrain);
end
indptest = (ytest==1);
indntest = ~indptest;
if(Ntest > 0)
indp = indptest;
indn = indntest;
Nneg = options.Nnegtest;
Npos = options.Npostest;
else
indp = indptrain;
indn = indntrain;
Nneg = options.Nnegtrain;
Npos = options.Npostrain;
end
alpham = 1;
betam = 1;
alphaold = 1;
betaold = 1;
K = 0;
m = 0;
cascade_old = options.cascade;
current_stage = length(options.m);
alpha0 = options.alpha0(current_stage);
beta0 = options.beta0(current_stage);
while( (m < options.maxwl_perstage) && ~((alpham < alpha0) && (betam < beta0)) )
m = m + 1;
wtrain = wtrain/sum(wtrain);
t1 = tic;
if(options.typefeat == 0)
if (options.algoboost == 0)
[hm , wnew] = haar_gentle_weaklearner(IItrain , ytrain , wtrain , options);
elseif (options.algoboost == 1)
[hm , wnew] = haar_ada_weaklearner(IItrain , ytrain , wtrain , options);
elseif (options.algoboost == 2)
[hm , wnew] = fast_haar_ada_weaklearner(IItrain , ytrain , wtrain , options);
elseif (options.algoboost == 3)
[hm , wnew] = haar_gentle_weaklearner_memory(Htrain , ytrain , wtrain , options);
else
[hm , wnew] = haar_ada_weaklearner_memory(Htrain , ytrain , wtrain , options);
end
elseif(options.typefeat == 1)
if (options.algoboost == 0)
[hm , wnew] = mblbp_gentle_weaklearner(Htrain , ytrain , wtrain , options);
elseif(options.algoboost == 1)
[hm , wnew] = mblbp_ada_weaklearner(Htrain , ytrain , wtrain , options);
end
end
t2 = toc(t1);
fprintf('Train weaklearner %d in %6.3f s\n' , m , t2);
drawnow
options.param = [options.param , hm];
options.cascade = [cascade_old , [m ; 0]];
if(Ntest > 0)
if(options.typefeat == 0)
fx = eval_haar(Xtest , options);
elseif(options.typefeat == 1)
fx = eval_mblbp(Xtest , options);
end
else
if(options.typefeat == 0)
fx = eval_haar(Xtrain , options);
elseif(options.typefeat == 1)
fx = eval_mblbp(Xtrain , options);
end
end
if(options.cascade_type == 0)
ufx = unique(fx);
co = 1;
currentthresh = ufx(co);
indfa = (fx(indn) >= currentthresh);
options.fa = sum(indfa);
currentalpha = options.fa/Nneg; %False acceptance rate
options.fr = sum(fx(indp) < currentthresh);
currentbeta = options.fr/Npos; %False rejection rate
threshold = currentthresh;
alpham = currentalpha;
betam = currentbeta;
while(currentbeta < beta0)
co = co+1;
threshold = currentthresh;
currentthresh = ufx(co);
alpham = currentalpha;
betam = currentbeta;
indfa = (fx(indn) >= currentthresh);
options.fa = sum(indfa);
currentalpha = options.fa/Nneg; %False acceptance rate
options.fr = sum(fx(indp) < currentthresh);
currentbeta = options.fr/Npos; %False rejection rate
end
elseif(options.cascade_type == 1)
threshold = 0;
indfa = (fx(indn) >= threshold);
options.fa = sum(indfa);
alpham = options.fa/Nneg; %False acceptance rate
options.fr = sum(fx(indp) < threshold);
betam = options.fr/Npos; %False rejection rate
end
if( (alpham == alphaold) || (betam == betaold))
K = K+1;
else
alphaold = alpham;
betaold = betam;
K = 0;
end
if(K >= options.maxK)
break;
end
wtrain = wnew;
options.indexF(hm(1)) = -1;
fprintf('stage %d/%d, m = %d, alpham = %5.4f\n' , current_stage , options.maxstage , m , alpham);
fprintf('stage %d/%d, m = %d, betam = %5.4f\n' , current_stage , options.maxstage , m , betam);
drawnow
end
if(options.cascade_type == 0)
Xfa = Xtest(: , : , indfa);
elseif(options.cascade_type == 1)
Xfa = Xtrain(: , : , indfa);
end
if(K < options.maxK)
options.m = [options.m , m];
options.thresholdperstage = [options.thresholdperstage , threshold];
options.alphaperstage = [options.alphaperstage , alpham];
options.betaperstage = [options.betaperstage , betam];
options.cascade(2,end) = threshold;
else
options.param = options.param(: , end-m+1);
end
