DLT（DeepLearningTracker）學習與代碼理解 (1)

 

持續更新中……

 

 

DLT是香港科大的Naiyan Wang和Dit-Yan Yeung提出的，文章和代碼見 http://winsty.net/dlt.html

相關論文是 Learning A Deep Compact Image Representation for Visual Tracking (NIPS2013')

 

論文本身並無太多亮點，可以說基本沒講啥。

大體思路是先用SDAE（Stacked Denoising AutoEncoder）對很多數據進行離線訓練，離線訓練的目的是提取特征。然后再用來在線跟蹤。跟蹤時使用粒子濾波來完成。

 

還是從研究代碼開始吧。

run_individual.m只是只是提供了一些圖像信息及參數，核心的函數在run_DLT.m里實現。

 

在這里必須吐槽一下MATLAB 2012a，如果系統語言為英文的話，居然沒法在M文件編輯器里用中文寫注釋。

 

 

function results=run_DLT(seq, res_path, bSaveImage) 
    addpath('affineUtility');
    addpath('drawUtility');
    addpath('imageUtility');
    addpath('NN');
    rand('state',0);  randn('state',0);
    
    %  如果opt未定義  
  9     if isfield(seq, 'opt')
        opt = seq.opt;
    else
        trackparam_DLT;
    end
    rect=seq.init_rect; % [水平左側 垂直上方 寬度 高度 0] 
    % 下面它會再蛋疼地改成 [水平中心 垂直中心 寬度 高度 0] 
    p = [rect(1)+rect(3)/2, rect(2)+rect(4)/2, rect(3), rect(4), 0];
    
    frame = imread(seq.s_frames{1});    % seq.s_frames 里放着所有圖像的路徑
    if size(frame,3)==3
        frame = double(rgb2gray(frame));   
    end
    
    scaleHeight = size(frame, 1) / opt.normalHeight;    % 讀入圖像的高度/標准圖像的高度，下同 24     scaleWidth = size(frame, 2) / opt.normalWidth;
    p(1) = p(1) / scaleWidth;   % 將GT中的目標位置參數按比例轉成標准圖像大小的對應值 26     p(3) = p(3) / scaleWidth;
    p(2) = p(2) / scaleHeight;
    p(4) = p(4) / scaleHeight;
    frame = imresize(frame, [opt.normalHeight, opt.normalWidth]); % 將讀入圖像進行比例放縮
    frame = double(frame) / 255;    
    
    
    % paramOld [ 模板水平中心 模板垂直中心, 寬度/標准模板寬度(32)
    %               0, 模板寬高比/標准模板寬高比(1:1), 0  ]
    paramOld = [p(1), p(2), p(3)/opt.tmplsize(2), p(5), p(4) /p(3) / (opt.tmplsize(1) / opt.tmplsize(2)), 0];
    param0 = affparam2mat(paramOld);    
    % param0 [ 模板水平中心 模板垂直中心, 寬度/標准模板寬度(32), 0, 高度/標准模板高度(32), 0 ]
   
    
    if ~exist('opt','var')  opt = [];  end
    if ~isfield(opt,'minopt')
      opt.minopt = optimset; opt.minopt.MaxIter = 25; opt.minopt.Display='off';
    end
    reportRes = [];
    
    % 處理模板（ 變換后） 
 48     tmpl.mean = warpimg(frame, param0, opt.tmplsize); % tmlsize = 32X32
    tmpl.basis = [];
    
    % 10個正樣本，在模板內取 
 52     for i = 1 : opt.maxbasis / 10
        tmpl.basis(:, (i - 1) * 10 + 1 : i * 10) = samplePos_DLT(frame, param0, opt.tmplsize);
    end
    % 100個負樣本（好像是在模板外隨機……待驗證） 
 56     p0 = paramOld(5);
    tmpl.basis(:, opt.maxbasis + 1 : 100 + opt.maxbasis) = sampleNeg(frame, param0, opt.tmplsize, 100, opt, 8);

    param.est = param0;
    param.lastUpdate = 1;

    wimgs = [];

    % draw initial track window
    drawopt = drawtrackresult([], 0, frame, tmpl, param, []);
    drawopt.showcondens = 0;  drawopt.thcondens = 1/opt.numsample;
    if (bSaveImage)
        imwrite(frame2im(getframe(gcf)),sprintf('%s0000.jpg',res_path));    
    end
    
    % track the sequence from frame 2 onward
    duration = 0; tic;
    if (exist('dispstr','var'))  dispstr='';  end
    L = [ones(opt.maxbasis, 1); (-1) * ones(100, 1)];
    
    % initialize NN and train
    nn = initDLT(tmpl, L);
    L = [];
    % add tmpl.basis to pos(:,11) cuz opt.maxbasis = 10
    pos = tmpl.basis(:, 1 : opt.maxbasis);
    pos(:, opt.maxbasis + 1) = tmpl.basis(:, 1);
    opts.numepochs = 5 ;
    
    % image iteration start
    for f = 1:size(seq.s_frames,1)  
      frame = imread(seq.s_frames{f});
      
      % if RGB ->gray
      if size(frame,3)==3
        frame = double(rgb2gray(frame));
      end  
      frame = imresize(frame, [opt.normalHeight, opt.normalWidth]);
      frame = double(frame) / 255;

      % do tracking
       param = estwarp_condens_DLT(frame, tmpl, param, opt, nn, f);

      % do update

      temp = warpimg(frame, param.est', opt.tmplsize);
      pos(:, mod(f - 1, opt.maxbasis) + 1) = temp(:);
      if  param.update
          opts.batchsize = 10;
          % Sample two set of negative samples at different range.
          neg = sampleNeg(frame, param.est', opt.tmplsize, 49, opt, 8);
          neg = [neg sampleNeg(frame, param.est', opt.tmplsize, 50, opt, 4)];
          nn = nntrain(nn, [pos neg]', [ones(opt.maxbasis + 1, 1); zeros(99, 1)], opts);
      end

      duration = duration + toc;
      
      res = affparam2geom(param.est);
      p(1) = round(res(1));
      p(2) = round(res(2)); 
      p(3) = round(res(3) * opt.tmplsize(2));
      p(4) = round(res(5) * (opt.tmplsize(1) / opt.tmplsize(2)) * p(3));
      p(5) = res(4);
      p(1) = p(1) * scaleWidth;
      p(3) = p(3) * scaleWidth;
      p(2) = p(2) * scaleHeight;
      p(4) = p(4) * scaleHeight;
      paramOld = [p(1), p(2), p(3)/opt.tmplsize(2), p(5), p(4) /p(3) / (opt.tmplsize(1) / opt.tmplsize(2)), 0];
      
      reportRes = [reportRes;  affparam2mat(paramOld)];
      
      tmpl.basis = [pos];
      drawopt = drawtrackresult(drawopt, f, frame, tmpl, param, []);
      if (bSaveImage)
          imwrite(frame2im(getframe(gcf)),sprintf('%s/%04d.jpg',res_path,f));
      end
      tic;
    end
    duration = duration + toc
    fprintf('%d frames took %.3f seconds : %.3fps\n',f,duration,f/duration);
    results.res=reportRes;
    results.type='ivtAff';
    results.tmplsize = opt.tmplsize;
    results.fps = f/duration;
end