组会

t-sne

相关性

SVM 线性核 RBM核 多项式核.

特征预处理：白化 归一化 减均值.

PCA之后白化效果会变好.

如果效果还是不好可以换换模型.

e-svm 模板经过缩放后，选取特征.选取你需要的HOG 大小.

检测时候，图像缩放不同尺度在提取特征,然后模板在不同的尺度的特征上滑动，问题出现了，这样做原因是 怕遗漏掉目标吗?相比较，不同尺度图像上sliding window然后提取特征，精度是否会提高？（但这样会带来的计算量是非常大的.）

>> a = [1 2 3;4 5 6]
a =
     1     2     3
     4     5     6
>> a(2)
     4

for i = 1:length(models)
  t = zeros(S(1),S(2),fsize);
  t(1:models{i}.model.hg_size(1),1:models{i}.model.hg_size(2),:) = ...
      models{i}.model.w;


  templates(:,:,:,i) = t;
  template_masks(:,:,:,i) = repmat(double(sum(t.^2,3)>0),[1 1 fsize]);


  if (~isempty(params.nnmode)) || ...
        (isfield(params,'wtype') && ...
         strcmp(params.wtype,'dfun')==1)
    x = zeros(S(1),S(2),fsize);
    x(1:models{i}.model.hg_size(1),1:models{i}.model.hg_size(2),:) = ...
        reshape(models{i}.model.x(:,1),models{i}.model.hg_size);
    templates_x(:,:,:,i) = x;


  end
end


%maskmat = repmat(template_masks,[1 1 1 fsize]);
%maskmat = permute(maskmat,[1 2 4 3]);
%templates_x  = templates_x .* maskmat;


sbin = models{1}.model.init_params.sbin;
t = get_pyramid(I, sbin, params);
resstruct.padder = t.padder;


pyr_N = cellfun(@(x)prod([size(x,1) size(x,2)]-S+1),t.hog);
sumN = sum(pyr_N);


X = zeros(S(1)*S(2)*fsize,sumN);
offsets = cell(length(t.hog), 1);
uus = cell(length(t.hog),1);
vvs = cell(length(t.hog),1);


counter = 1;
for i = 1:length(t.hog)
  s = size(t.hog{i});
  NW = s(1)*s(2);
  ppp = reshape(1:NW,s(1),s(2));
  curf = reshape(t.hog{i},[],fsize);
  b = im2col(ppp,[S(1) S(2)]);%准备在特征上滑动窗口，每个窗口内像素索引值作为b的一列


  offsets{i} = b(1,:);%在特征上滑动窗口，记录每个窗口左上角索引值
  offsets{i}(end+1,:) = i;
  
  for j = 1:size(b,2)
   X(:,counter) = reshape (curf(b(:,j),:),[],1);%在特征上滑动窗口
   counter = counter + 1;
  end
  
  [uus{i},vvs{i}] = ind2sub(s,offsets{i}(1,:));%每个窗口左上角索引值转变成x,y坐标
end


offsets = cat(2,offsets{:});%合并所有窗口左上角的索引值


uus = cat(2,uus{:});
vvs = cat(2,vvs{:});


% m.model.w = zeros(S(1),S(2),fsize);
% m.model.b = 0;
% temp_params = params;
% temp_params.detect_save_features = 1;
% temp_params.detect_exemplar_nms_os_threshold = 1.0;
% temp_params.max_models_before_block_method = 1;
% temp_params.detect_max_windows_per_exemplar = 28000;


% [rs] = esvm_detect(I, {m}, temp_params);
% X2=cat(2,rs.xs{1}{:});
% bbs2 = rs.bbs{1};




exemplar_matrix = reshape(templates,[],size(templates,4));


if isfield(params,'wtype') && ...
      strcmp(params.wtype,'dfun')==1
  W = exemplar_matrix;
  U = reshape(templates_x,[],length(models));
  r2 = repmat(sum(W.*(U.^2),1)',1,size(X,2));
  r =  (W'*(X.^2) - 2*(W.*U)'*X + r2);
  r = bsxfun(@minus, r, bs);
elseif isempty(params.nnmode)%得分减bs，取最小误差的窗口
  %nnmode 0: Apply linear classifiers by performing one large matrix
  %multiplication and subtract bias
  r = exemplar_matrix' * X;
  r = bsxfun(@minus, r, bs);
elseif strcmp(params.nnmode,'normalizedhog') == 1
  r = exemplar_matrix' * X;
elseif strcmp(params.nnmode,'nndfun') == 1
  %Do euclidean distance (but only over the regions corresponding
  %to the in-mask (non-padded) regions
  W = reshape(template_masks,[],length(models));
  W = W / 100;
  U = reshape(templates_x,[],length(models));
  r2 = repmat(sum(W.*(U.^2),1)',1,size(X,2));
  r = - (W'*(X.^2) - 2*(W.*U)'*X + r2);
else
  error('invalid nnmode=%s\n',params.nnmode);
end


resstruct.bbs = cell(N,1);
resstruct.xs = cell(N,1);


for exid = 1:N


  goods = find(r(exid,:) >= params.detect_keep_threshold);
  
  if isempty(goods)
    continue
  end
  
  [sorted_scores,bb] = ...
      psort(-r(exid,goods)',...
            min(params.detect_max_windows_per_exemplar, ...%取前K个误差最小的窗体
                length(goods)));
  bb = goods(bb);


  sorted_scores = -sorted_scores';


  resstruct.xs{exid} = X(:,bb);
  
  levels = offsets(2,bb);
  scales = t.scales(levels);
  curuus = uus(bb);
  curvvs = vvs(bb);
  o = [curuus' curvvs'] - t.padder;


  bbs = ([o(:,2) o(:,1) o(:,2)+size(ws{exid},2) ...
           o(:,1)+size(ws{exid},1)] - 1) .* ...
             repmat(sbin./scales',1,4) + 1 + repmat([0 0 -1 ...
                    -1],length(scales),1);%% 有待了解.
  
  bbs(:,5:12) = 0;
  bbs(:,5) = (1:size(bbs,1));
  bbs(:,6) = exid;
  bbs(:,8) = scales;
  bbs(:,9) = uus(bb);
  bbs(:,10) = vvs(bb);
  bbs(:,12) = sorted_scores;
  
  if (params.detect_add_flip == 1)
    bbs = flip_box(bbs,t.size);
    bbs(:,7) = 1;
  end
  
  resstruct.bbs{exid} = bbs;
end




if params.detect_save_features == 0
  resstruct.xs = cell(N,1);
end
%fprintf(1,'\n');