原文链接 http://wangcaiyong.com/2015/08/15/sdm2/
注：以下为加速网络访问所做的原文缓存，经过重新格式化，可能存在格式方面的问题，或偶有遗漏信息，请以原文为准。
训练阶段我们采用了分批处理，可以优化部分内存。
原先我们的代码使用一次性载入数据，然后开始若干次迭代，直至收敛。这样保存image与shape的数据矩阵Data就一直占用内存，但实际上，数据集的image与shape 的作用仅仅是计算$\Delta X$和$\Phi$，两变量的定义参考《 Supervised Descent Method and its Applications to Face Alignment》，其实也就是init shape与true shape的差值以及init shape的特征向量。计算完了，Data就没用了。于是我们设想分批处理，每次迭代载入一次数据，用完了就clear。这样就需要载入多次，可能时间上会浪费的比较多，但还好，可以承受。而且每次载入数据的时候，也使用分批处理，一次只载入10幅图片的Image/shape，算完了就clear。 <!--more-->
程序中，我们为了增加样本的数量也使用了flip image操作，你可以选择flip or not flip,可以在setup.m中设置
options.flipFlag = 1; % the flag of flipping。
如下为分批处理的代码：
%确定分批的序号
if (isfield(options, 'batch_size'))%批量处理
batch_fid = cell(ceil(nData / (options.batch_size*basize)), 1);% batch_index为批量处理的索引号， batch_fid为批量处理的原图片序号
for bb = 1:length(batch_fid)
batch_fid{bb} = ...
(bb-1)*options.batch_size+1: ...
min(bb*options.batch_size,length(imlist));
end
end
ndata_i=0;%nData的遍历序号
for idata = 1 : length(batch_fid)
for jdata=1:length( batch_fid{idata})
%% the information of i-th image
%disp(Data(idata).img);
disp(['Stage: ' num2str(options.current_cascade) ' - Image: ' num2str(batch_fid{idata}(jdata))]);
Data= load_single_data2 ( imgDir, ptsDir, batch_fid{idata}(jdata),options );
for kdata=1:basize
ndata_i=ndata_i+1;
img = Data.img_gray;
shape = Data.shape_gt;
%% fipping data,左右翻转图像，用于产生更多的训练数据
if kdata==2
if size(img,3) > 1
img_gray = fliplr(rgb2gray(uint8(img)));
else
img_gray = fliplr(img);
end
clear img;
img = img_gray;
clear img_gray;
shape = flipshape(shape);
shape(:,1) = size(img,2) - shape(:, 1);
if 0
figure(1); imshow(img); hold on;
draw_shape(shape(:,1),...
shape(:,2),'y');
hold off;
pause;
end
end
%% if the first cascade
if ( current_cascade == 1 )
%% if detect face using viola opencv
%boxes = detect_face( img , options );
%% if using ground-truth
bbox = [];
%% predict the face box
if isempty(bbox)
%% if using ground-truth
bbox = getbbox(shape);
end
%% randomize n positions for initial shapes
[rbbox] = random_init_position( ...
bbox, DataVariation, n_init_randoms,options );%获得n个随机的真实人脸框，模拟人脸识别detector
%% iterations of n initial points
for ir = 1 : n_init_randoms
%% get random positions and inital shape indexs
cbbox = rbbox(ir,:);
init_shape = resetshape(cbbox, MeanShape2);%将平均人脸（而不是初始化人脸）对齐到每个算出的真实人脸框上
if 0
figure(1); imshow(img); hold on;
rectangle('Position', cbbox, 'EdgeColor', 'y');
draw_shape(init_shape(:,1),...
init_shape(:,2),'y');
hold on;
rectangle('Position', bbox, 'EdgeColor', 'r');
draw_shape(shape(:,1),...
shape(:,2),'r');
hold off;
pause;
end
%% scale coarse to fine %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
cropIm_scale = imresize(img,current_scale);
init_shape = init_shape * current_scale;
true_shape = shape * current_scale;
if 0
figure(1); imshow(cropIm_scale); hold on;
draw_shape(init_shape(:,1), init_shape(:,2),'g');
draw_shape(true_shape(:,1), true_shape(:,2),'r');
hold off;
% pause;
end
storage_bbox((ndata_i-1)*n_init_randoms+ir,:) = ...
getbbox(init_shape);
%% compute the descriptors and delta_shape %%%%%%%%%%%%%%%%%%%%
% storing the initial shape
storage_init_shape((ndata_i-1)*n_init_randoms+ir,:) = ...
shape_2_vec(init_shape);
% storing the the descriptors
tmp = local_descriptors( cropIm_scale, ...
init_shape,...
desc_size, desc_bins, options );
storage_init_desc((ndata_i-1)*n_init_randoms+ir,:) = tmp(:);
% storing delta shape
tmp_del = init_shape - true_shape;
shape_residual = bsxfun(@rdivide, tmp_del, ...
storage_bbox((ndata_i-1)*n_init_randoms+ir,3:4));
storage_del_shape((ndata_i-1)*n_init_randoms+ir,:) = ...
shape_2_vec(shape_residual);
storage_gt_shape((ndata_i-1)*n_init_randoms+ir,:) = ...
shape_2_vec(shape);
% storage_image{(ndata_i-1)*n_init_randoms+ir}=cropIm_scale;%补充项
end
else
% for higher cascaded levels
for ir = 1 : n_init_randoms
init_shape = vec_2_shape(new_init_shape((ndata_i-1)*n_init_randoms+ir,:)');
%% scale coarse to fine %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% scale coarse to fine %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
cropIm_scale = imresize(img,current_scale);
init_shape = init_shape * current_scale;
true_shape = shape * current_scale;
%% compute the descriptors and delta_shape %%%%%%%%%%%%%%%%%%%%
if 0
figure(1); imshow(cropIm_scale); hold on;
draw_shape(init_shape(:,1), init_shape(:,2),'g');
draw_shape(true_shape(:,1), true_shape(:,2),'r');
hold off;
pause;
end
storage_bbox((ndata_i-1)*n_init_randoms+ir,:) = ...
getbbox(init_shape);
%% compute the descriptors and delta_shape %%%%%%%%%%%%%%%%%%%%
% storing the initial shape
storage_init_shape((ndata_i-1)*n_init_randoms+ir,:) = ...
shape_2_vec(init_shape);
% storing the the descriptors
tmp = local_descriptors( cropIm_scale, ...
init_shape,...
desc_size, desc_bins, options );
storage_init_desc((ndata_i-1)*n_init_randoms+ir,:) = tmp(:);
% storing delta shape
tmp_del = init_shape - true_shape;
shape_residual = bsxfun(@rdivide, tmp_del, ...
storage_bbox((ndata_i-1)*n_init_randoms+ir,3:4));
storage_del_shape((ndata_i-1)*n_init_randoms+ir,:) = ...
shape_2_vec(shape_residual);
storage_gt_shape((ndata_i-1)*n_init_randoms+ir,:) = ...
shape_2_vec(shape);
% storage_image{(ndata_i-1)*n_init_randoms+ir}=cropIm_scale;%补充项
end
end
clear img;
clear cropIm_scale;
clear shape;
end
clear Data;
end
end