multimodefacerecognition.m

It is for Face Recognition

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% function kfdFaceRecognition()
%   z.li, 07-14-2004
%   multi mode face recogntion: eig, fisher, kernel-fisher face based methods
% function dependency:
%   - mixedClassify()
%   - getATTFaceProj()
% input:
%   fm              - face model: dim x n
%   qId, qPose      - an unknown face
%   ids             - known persons
%   poses           - known poses
%   wt              - distance metric weights: dim x 1
%   R               - local neighbourhood radius
% output:
%   recId, nnId, fdaId, knlId, recErr, fdaErr, knlErr, m
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%function [recId, nnId, fdaId, knlId, recErr, fdaErr, knlErr, m]=multiModeFaceRecognition(fm, ids, poses, qId, qPose,wt, R)
function [recId, nnId, fdaId, knlId, recErr, fdaErr, knlErr, m]=multiModeFaceRecognition(fm, ids, poses, qId, qPose, wt, R)
dbg = 'n';
if dbg=='y'
    % 18x16 lower face, 12-d
    Fm1 = load('prjATTUpperFaceS04T08L08W18H16K16.dat');
    % 14x18 lower face, 12-d
    Fm2 = load('prjATTLowerFaceS01T-06L14W18H12K12.dat');
    % 21x28 full face,  20-d
    Fm3 = load('prjATTFullFaceModel4.dat');    

   fm = Fm1; 
   [dim, nFace]=size(fm);   
   ids = 11:30; poses = [1 2 3 4 6 7];
   ids = [11:40]; poses = [1 3 4 6 7 9]; xPoses = [2 5 8 10];
   qId =36; qPose = 2;
   wt = ones(dim, 1); R=860; 
end

%const
fitLocalFDA = 'y';
minPoseCnt = 3;  minLocalSampleRatio = 0.2; defaultNNErr = 0.005;

nPerson     = length(ids);
nPose       = length(poses);

% face model
[dim, nFace]=size(fm);
% distance metric weight
W=eye(dim);
for k=1:dim  
  W(k,k)=wt(k);
end

% get unknown face: dim x 1, col vec
xFace = getAttFaceProj(fm, qId, qPose);
if isnan(xFace(1))== 1      % not exist
   recId=-2;  nnId=-2; fdaId=-2; knlId=-2; 
   recErr=inf; fdaErr=inf; knlErr=inf; m=0;
   fprintf('x');
   return;
end

% build local model: X: m x dim, Y: mx1
m=0;
for j=1:length(ids)
  thisId = ids(j);
  % check if this person has enough sample, if so add to model
  kCnt=0; 
  dist=zeros(length(poses), 1);
  for k=1:length(poses)         % for a person
     thisFace=getAttFaceProj(fm, thisId, poses(k)); % col vec
     diff = thisFace-xFace;
     dist(k) = sqrt(diff'*W*diff);
     if dist(k) < R
        kCnt=kCnt+1;
     end
  end % poses

  if kCnt >= minPoseCnt     % enough sample in local model
     % add to the local model
     for k=1:length(poses)  
        if dist(k) < R
            m=m+1;
            thisFace=getAttFaceProj(fm, thisId, poses(k)); % col vec
            X(m,:)=[thisFace']; Y(m)=thisId;
        end       
     end
  end %of enough sample
end % ids

% check the local model
if m <= 0  
  recId=-3; nnId=-3; fdaId=-3; knlId=-3;
  recErr=inf; fdaErr=inf; knlErr=inf; m=0;
  fprintf('m');
  return;
end

% check the NN result
dv=[];
for k=1:m
  dv(k)=norm(xFace'-X(k,:), 2);
end
[minDist, idx]=min(dv);
nnId = Y(idx);

% fit local model
if fitLocalFDA == 'y'
    nDim = length(unique(Y));
    [ev, A]=lda([X Y'], min(nDim-1, dim));
    %projection
    for j=1:length(Y)
       X2(j,:)=X(j,:)*A;
    end
    X = X2;
    xFace=(xFace'*A)';
end

[fdaId, fdaErr]=mixedClassify(xFace', X, Y, 'linear');
if isinf(fdaErr)==1
  fdaId = -1;
end

% check the kernel FDA result
[knlId, knlErr]=mixedClassify(xFace', X, Y, 'quadratic');
if isinf(knlErr)==1
  knlId = -1;
end

% combine results
if fdaId == -1 | (m/(nPerson*nPose)) <= minLocalSampleRatio
   recId = nnId; recErr = defaultNNErr; fprintf('n');
else
   recId = fdaId; recErr = fdaErr; fprintf('f');
end