跑benchmark遇到的那些坑
- 如果在运行“perfPlot.m“文件时提示” 未定义函数或变量 'calcRectInt' “，只要把“restEval”文件夹addpath到就好了perfPlots
- “结构体内容引用自非结构体对象” 数据集的问题，没配对上，把出问题的那条数据集注释了就行
- （这条是抄的）util/configTrcakers.m文件，打开，这里面是所有跟踪算法名字的数组，会在main_running.m和perfPlot.m中被调用到。因为results文件夹中已经包含已有跟踪算法的测试结果，所以其实没必要再重新跑一遍所有算法，时间太长。故而直接将configTrackers.m文件中的所有注释掉，改成如下三行代码即可：（就是说让这个文件夹里只有你自己写的跟踪算法，这样一会执行main_running.m的时候，就不用在测试集上挨个跑已有跟踪算法了）
function trackers=configTrackers
trackersNEW={struct('name','LCSVM','namePaper','LCSVM')};
trackers=trackersNEW;
————————————————
版权声明：本文为CSDN博主「劲草浅躬行」的原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/lcb_coconut/article/details/76512707 -
错误使用 vertcat 串联的矩阵的维度不一致
没事儿 改去吧 你数据集有问题 重新复制粘贴一行行改地址去吧
- 最后跑通的main_running如下
-
close all clear clc warning off all; addpath('./util'); addpath(('E:\code\vlfeat-0.9.21-bin\vlfeat-0.9.21\toolbox')); vl_setup addpath(('./rstEval')); addpath(['./trackers/VIVID_Tracker']) seqs=configSeqs; trackers=configTrackers; shiftTypeSet = {'left','right','up','down','topLeft','topRight','bottomLeft','bottomRight','scale_8','scale_9','scale_11','scale_12'}; evalType='TRE'; %'OPE','SRE','TRE' diary(['./tmp/' evalType '.txt']); numSeq=length(seqs); numTrk=length(trackers); finalPath = ['./results/results_' evalType '_CVPR13/']; if ~exist(finalPath,'dir') mkdir(finalPath); end tmpRes_path = ['./tmp/' evalType '/']; bSaveImage=0; if ~exist(tmpRes_path,'dir') mkdir(tmpRes_path); end pathAnno = './anno/'; for idxSeq=1:length(seqs) s = seqs{idxSeq}; % if ~strcmp(s.name, 'coke') % continue; % end s.len = s.endFrame - s.startFrame + 1; s.s_frames = cell(s.len,1); nz = strcat('%0',num2str(s.nz),'d'); %number of zeros in the name of image for i=1:s.len image_no = s.startFrame + (i-1); id = sprintf(nz,image_no); s.s_frames{i} = strcat(s.path,id,'.',s.ext); end img = imread(s.s_frames{1}); [imgH,imgW,ch]=size(img); rect_anno = dlmread([pathAnno s.name '.txt']); numSeg = 20; [subSeqs, subAnno]=splitSeqTRE(s,numSeg,rect_anno); switch evalType case 'SRE' subS = subSeqs{1}; subA = subAnno{1}; subSeqs=[]; subAnno=[]; r=subS.init_rect; for i=1:length(shiftTypeSet) subSeqs{i} = subS; shiftType = shiftTypeSet{i}; subSeqs{i}.init_rect=shiftInitBB(subS.init_rect,shiftType,imgH,imgW); subSeqs{i}.shiftType = shiftType; subAnno{i} = subA; end case 'OPE' subS = subSeqs{1}; subSeqs=[]; subSeqs{1} = subS; subA = subAnno{1}; subAnno=[]; subAnno{1} = subA; otherwise end for idxTrk=1:numTrk t = trackers{idxTrk}; % if ~strcmp(t.name, 'LSK') % continue; % end % validate the results if exist([finalPath s.name '_' t.name '.mat']) load([finalPath s.name '_' t.name '.mat']); bfail=checkResult(results, subAnno); if bfail disp([s.name ' ' t.name]); end continue; end switch t.name case {'VTD','VTS'} continue; end results = []; for idx=1:length(subSeqs) disp([num2str(idxTrk) '_' t.name ', ' num2str(idxSeq) '_' s.name ': ' num2str(idx) '/' num2str(length(subSeqs))]) rp = [tmpRes_path s.name '_' t.name '_' num2str(idx) '/']; if bSaveImage&~exist(rp,'dir') mkdir(rp); end subS = subSeqs{idx}; subS.name = [subS.name '_' num2str(idx)]; % subS.s_frames = subS.s_frames(1:20); % subS.len=20; % subS.endFrame=subS.startFrame+subS.len-1; funcName = ['res=run_' t.name '(subS, rp, bSaveImage);']; try switch t.name case {'VR','TM','RS','PD','MS'} otherwise cd(['./trackers/' t.name]); addpath(genpath('./')) end eval(funcName); switch t.name case {'VR','TM','RS','PD','MS'} otherwise rmpath(genpath('./')) cd('../../'); end if isempty(res) results = []; break; end catch err disp('error'); rmpath(genpath('./')) cd('../../'); res=[]; continue; end res.len = subS.len; res.annoBegin = subS.annoBegin; res.startFrame = subS.startFrame; switch evalType case 'SRE' res.shiftType = shiftTypeSet{idx}; end results{idx} = res; end save([finalPath s.name '_' t.name '.mat'], 'results'); end end figure t=clock; t=uint8(t(2:end)); disp([num2str(t(1)) '/' num2str(t(2)) ' ' num2str(t(3)) ':' num2str(t(4)) ':' num2str(t(5))]);最后跑通的perfPlot如下
-
clear close all; clc addpath('./util'); addpath(('./rstEval')); attPath = '.\anno\att\'; % The folder that contains the annotation files for sequence attributes attName={'illumination variation' 'out-of-plane rotation' 'scale variation' 'occlusion' 'deformation' 'motion blur' 'fast motion' 'in-plane rotation' 'out of view' 'background clutter' 'low resolution'}; attFigName={'illumination_variations' 'out-of-plane_rotation' 'scale_variations' 'occlusions' 'deformation' 'blur' 'abrupt_motion' 'in-plane_rotation' 'out-of-view' 'background_clutter' 'low_resolution'}; plotDrawStyleAll={ struct('color',[1,0,0],'lineStyle','-'),... struct('color',[0,1,0],'lineStyle','-'),... struct('color',[0,0,1],'lineStyle','-'),... struct('color',[0,0,0],'lineStyle','-'),...% struct('color',[1,1,0],'lineStyle','-'),...%yellow struct('color',[1,0,1],'lineStyle','-'),...%pink struct('color',[0,1,1],'lineStyle','-'),... struct('color',[0.5,0.5,0.5],'lineStyle','-'),...%gray-25% struct('color',[136,0,21]/255,'lineStyle','-'),...%dark red struct('color',[255,127,39]/255,'lineStyle','-'),...%orange struct('color',[0,162,232]/255,'lineStyle','-'),...%Turquoise struct('color',[163,73,164]/255,'lineStyle','-'),...%purple %%%%%%%%%%%%%%%%%%%% struct('color',[1,0,0],'lineStyle','--'),... struct('color',[0,1,0],'lineStyle','--'),... struct('color',[0,0,1],'lineStyle','--'),... struct('color',[0,0,0],'lineStyle','--'),...% struct('color',[1,1,0],'lineStyle','--'),...%yellow struct('color',[1,0,1],'lineStyle','--'),...%pink struct('color',[0,1,1],'lineStyle','--'),... struct('color',[0.5,0.5,0.5],'lineStyle','--'),...%gray-25% struct('color',[136,0,21]/255,'lineStyle','--'),...%dark red struct('color',[255,127,39]/255,'lineStyle','--'),...%orange struct('color',[0,162,232]/255,'lineStyle','--'),...%Turquoise struct('color',[163,73,164]/255,'lineStyle','--'),...%purple %%%%%%%%%%%%%%%%%%% struct('color',[1,0,0],'lineStyle','-.'),... struct('color',[0,1,0],'lineStyle','-.'),... struct('color',[0,0,1],'lineStyle','-.'),... struct('color',[0,0,0],'lineStyle','-.'),...% struct('color',[1,1,0],'lineStyle',':'),...%yellow struct('color',[1,0,1],'lineStyle','-.'),...%pink struct('color',[0,1,1],'lineStyle','-.'),... struct('color',[0.5,0.5,0.5],'lineStyle','-.'),...%gray-25% struct('color',[136,0,21]/255,'lineStyle','-.'),...%dark red struct('color',[255,127,39]/255,'lineStyle','-.'),...%orange struct('color',[0,162,232]/255,'lineStyle','-.'),...%Turquoise struct('color',[163,73,164]/255,'lineStyle','-.'),...%purple }; plotDrawStyle10={ struct('color',[1,0,0],'lineStyle','-'),... struct('color',[0,1,0],'lineStyle','--'),... struct('color',[0,0,1],'lineStyle',':'),... struct('color',[0,0,0],'lineStyle','-'),...% struct('color',[1,1,0],'lineStyle','-'),...%yellow struct('color',[1,0,1],'lineStyle','--'),...%pink struct('color',[0,1,1],'lineStyle',':'),... struct('color',[0.5,0.5,0.5],'lineStyle','-'),...%gray-25% struct('color',[136,0,21]/255,'lineStyle','--'),...%dark red struct('color',[255,127,39]/255,'lineStyle',':'),...%orange struct('color',[0,162,232]/255,'lineStyle','-'),...%Turquoise }; seqs=configSeqs; trackers=configTrackers; % seqs = seqs(1:10); % trackers = trackers(1:10); numSeq=length(seqs); numTrk=length(trackers); nameTrkAll=cell(numTrk,1); for idxTrk=1:numTrk t = trackers{idxTrk}; nameTrkAll{idxTrk}=t.namePaper; end nameSeqAll=cell(numSeq,1); numAllSeq=zeros(numSeq,1); att=[]; for idxSeq=1:numSeq s = seqs{idxSeq}; nameSeqAll{idxSeq}=s.name; s.len = s.endFrame - s.startFrame + 1; numAllSeq(idxSeq) = s.len; att(idxSeq,:)=load([attPath s.name '.txt']); end attNum = size(att,2); figPath = '.\figs\overall\'; perfMatPath = '.\perfMat\overall\'; if ~exist(figPath,'dir') mkdir(figPath); end metricTypeSet = {'error', 'overlap'}; % evalTypeSet = {'SRE', 'TRE', 'OPE'}; evalTypeSet ='TRE';%我把这儿改了！！！！！！！ rankingType = 'AUC';%AUC, threshod rankingType = 'threshold';%AUC, threshold rankNum = 10;%number of plots to show if rankNum == 10 plotDrawStyle=plotDrawStyle10; else plotDrawStyle=plotDrawStyleAll; end thresholdSetOverlap = 0:0.05:1; thresholdSetError = 0:50; for i=1:length(metricTypeSet) metricType = metricTypeSet{i};%error,overlap switch metricType case 'overlap' thresholdSet = thresholdSetOverlap; rankIdx = 11; xLabelName = 'Overlap threshold'; yLabelName = 'Success rate'; case 'error' thresholdSet = thresholdSetError; rankIdx = 21; xLabelName = 'Location error threshold'; yLabelName = 'Precision'; end if strcmp(metricType,'error')&strcmp(rankingType,'AUC') continue; end tNum = length(thresholdSet); % for j=1:length(evalTypeSet) evalType = evalTypeSet;%SRE, TRE, OPE plotType = [metricType '_' evalType]; switch metricType case 'overlap' titleName = ['Success plots of ' evalType]; case 'error' titleName = ['Precision plots of ' evalType]; end dataName = [perfMatPath 'aveSuccessRatePlot_' num2str(numTrk) 'alg_' plotType '.mat']; % If the performance Mat file, dataName, does not exist, it will call % genPerfMat to generate the file. if ~exist(dataName) genPerfMat(seqs, trackers, evalType, nameTrkAll, perfMatPath); end load(dataName); numTrk = size(aveSuccessRatePlot,1); if rankNum > numTrk | rankNum <0 rankNum = numTrk; end figName= [figPath 'quality_plot_' plotType '_' rankingType]; idxSeqSet = 1:length(seqs); %% % % PREFORMATTED % TEXT % % draw and save the overall performance plot plotDrawSave(numTrk,plotDrawStyle,aveSuccessRatePlot,idxSeqSet,rankNum,rankingType,rankIdx,nameTrkAll,thresholdSet,titleName, xLabelName,yLabelName,figName,metricType); % draw and save the performance plot for each attribute attTrld = 0; for attIdx=1:attNum idxSeqSet=find(att(:,attIdx)>attTrld); if length(idxSeqSet) < 2 continue; end disp([attName{attIdx} ' ' num2str(length(idxSeqSet))]) figName=[figPath attFigName{attIdx} '_' plotType '_' rankingType]; titleName = ['Plots of ' evalType ': ' attName{attIdx} ' (' num2str(length(idxSeqSet)) ')']; switch metricType case 'overlap' titleName = ['Success plots of ' evalType ' - ' attName{attIdx} ' (' num2str(length(idxSeqSet)) ')']; case 'error' titleName = ['Precision plots of ' evalType ' - ' attName{attIdx} ' (' num2str(length(idxSeqSet)) ')']; end plotDrawSave(numTrk,plotDrawStyle,aveSuccessRatePlot,idxSeqSet,rankNum,rankingType,rankIdx,nameTrkAll,thresholdSet,titleName, xLabelName,yLabelName,figName,metricType); end end % end修改的跟踪代码Dsiam如下
-
function results= run_main(seq,res_path,bSaveIamge) %% % This code is a demo for DSiam, a online updated deep tracker for fast tracking % If you use this code,please cite: % Qing Guo, Wei Feng, Ce Zhou, Rui Huang, Liang Wan, Song Wang. % Learning Dynamic Siamese Network for Visual Object Tracking. In ICCV 2017. % % Qing Guo, 2017. %% %===================================% % Path initial setting %===================================% netname = 'siamfc'; % '1res' denotes the multi-layer DSiam (DSiamM in paper) and uses two layers for tracking % '0res' denotes the single-layer DSiam (DSiam in paper) and uses the last layer for tracking nettype = '0res'; rootdir = '../DSIAM/'; addpath(genpath([rootdir,'/matconvnet/matlab/'])); vl_setupnn ; addpath([rootdir,'utils']); addpath([rootdir,'models']); %===================================% % Data initialization %===================================% s_frames = seq.s_frames; initrect = seq.init_rect; % read img imgs = vl_imreadjpeg(s_frames,'numThreads', 12); img = imgs{1}; if size(img,3)==1%表示图像是灰度图 如果==3 为RGB彩色图 img = repmat(img,1,1,3);% x = repmat(img,[1,1,3]);%将单通道图片转换为三通道图片 end nFrames = numel(s_frames);%输出s_frames像素个数 state = obj_initialize(img, initrect);%这里是下面的obj_initialize函数 需要参数图像/初始矩形 state.seq.frame = 1; state.seq.nFrames = nFrames; state.seq.seqname = seq.name; state.seq.time = 0; state.seq.isfail = 0; state.seq.scores = zeros(nFrames); state.seq.frame = 1; %帧数序列 state.seq.scores(1) = 10; % default if nargin<4 %输入参数个数小于4 netname = 'vgg19'; nettype = '1res'; end [state, opts]... = fcnet_init(img, state); %fcnet_init函数完成网络孪生网络的初始化和相关参数设置等工作 %fcnet_init函数在utils文件夹中 res=[initrect]; duration = 0; for it = 2:nFrames %从第二帧开始 state.seq.frame = it; fprintf('Processing frame %d/%d\n... ', state.seq.frame, nFrames); % ********************************** % VOT: Get next frame % ********************************** img = imgs{it}; if size(img,3)==1 img = repmat(img,1,1,3); end state = fcn_update(state, img,opts); initstate = [state.obj.targetLoc]; duration = state.seq.time; res = [res; initstate]; %================================== %Display result %================================== % if bSaveImg %-------------------------------Show the tracking results imshow(uint8(img)); rectangle('Position',initstate,'LineWidth',1,'EdgeColor','r'); hold on; text(5, 18, strcat('#',num2str(it)), 'Color','y', 'FontWeight','bold', 'FontSize',20); set(gca,'position',[0 0 1 1]); pause(0.01); hold off; % saveas(gcf,[res_path num2str(i) '.jpg']) % imwrite(frame2im(getframe(gcf)),[res_path num2str(it) % '.jpg']);这两句是每帧截图的 % end end results.res=res; results.type='rect'; results.fps=(seq.len)/duration; disp(['fps: ' num2str(results.fps)]) end function [state] = obj_initialize(I, region, varargin) gray = double(I(:,:,1)); [height, width] = size(gray); % If the provided region is a polygon 如果提供的矩阵是多边形... if numel(region) > 4 x1 = round(min(region(1:2:end))); x2 = round(max(region(1:2:end))); y1 = round(min(region(2:2:end))); y2 = round(max(region(2:2:end))); region = round([x1, y1, x2 - x1, y2 - y1]); else region = round([round(region(1)), round(region(2)), ... round(region(1) + region(3)) - round(region(1)), ... round(region(2) + region(4)) - round(region(2))]); end; %画框框 x1 = max(0, region(1)); y1 = max(0, region(2)); x2 = min(width-1, region(1) + region(3) - 1); y2 = min(height-1, region(2) + region(4) - 1); %那这一块是干嘛的呢？ state.obj.pos = [y1 + y2 + 1, x1 + x2 + 1] / 2; state.obj.targetsz = [y2-y1+1,x2-x1+1]; state.obj.base_targetsz = [y2-y1+1,x2-x1+1]; state.obj.targetLoc = [x1, y1, state.obj.targetsz([2,1])]; state.obj.change_alphaf = []; state.obj.change_featf = []; end function state= fcn_update(state,img,opts) %load state targetsz = state.obj.targetsz.*opts.targetszrate; pos = state.obj.pos; s_x = state.obj.s_x; corrfeat = state.obj.corrfeat; isfail = state.seq.isfail; net_conv = state.net.net_conv; net_obj = state.net.net_obj; %load params instanceSize = opts.instanceSize; window = opts.window; min_s_x = opts.min_s_x; max_s_x = opts.max_s_x; avgChans = opts.avgChans; scales = opts.scales; scaledInstance = s_x .* scales; scaledTarget = [targetsz(1) .* scales; targetsz(2) .* scales]; tic; % extract scaled crops for search region x at previous target position 在前一目标位置提取搜索区域X的缩放作物 x_crops = make_scale_pyramid(img, pos, scaledInstance, instanceSize, avgChans,opts);%, opts,saliency_map);显著性地图 %代码提取了三种尺度下的以目标为中心的图像patch块。然后使用孪生网络中用于跟踪的网络（网络结构如图1所示）进行一次前向传播，得到新目标位置等。到此完成跟踪当前帧，只需要不断迭代这一步骤即可。 % evaluate the offline-trained network for exemplar x features评估示例性X特征的离线训练网络 [newTargetPosition, newScale, score,responseMap,scorePos] = tracker_eval(net_conv,round(s_x), ... corrfeat, x_crops, pos, window, opts); pos = gather(newTargetPosition); if opts.isupdate % score if score >0 wc_z = targetsz(2) + opts.contextAmount*sum(targetsz); hc_z = targetsz(1) + opts.contextAmount*sum(targetsz); s_z = sqrt(wc_z*hc_z); [z_crop, ~] = get_subwindow_tracking(img, pos, ... [opts.exemplarSize opts.exemplarSize], [round(s_z) round(s_z)], opts.avgChans,opts.averageImage); z_crop = gpuArray(single(z_crop)); net_obj.eval({opts.netobj_input, z_crop}); state.seq.scores(state.seq.frame) = score; tcorrfeat{1} = net_obj.vars(opts.obj_feat_id(1)).value; tcorrfeat{2} = net_obj.vars(opts.obj_feat_id(2)).value; % updating the target variation transformation更新目标变化转换 if opts.vartransform net_conv = update_v(net_conv,corrfeat,tcorrfeat,opts); end % updating the background suppression transformation if opts.backsupression [x_back(:,:,:,1), ~] = get_subwindow_tracking(gather(img), pos,... [instanceSize instanceSize], [round(scaledInstance(newScale)) round(scaledInstance(newScale))], avgChans); x_back(:,:,:,2) = x_back(:,:,:,1).* opts.saliency_window;%state.obj.x_crop;% net_obj.eval({opts.netobj_input, gpuArray(x_back)}); tcorrfeat{1} = net_obj.vars(opts.obj_feat_id(1)).value; tcorrfeat{2} = net_obj.vars(opts.obj_feat_id(2)).value; net_conv = update_w(net_conv,tcorrfeat,opts); end % scale damping and saturation比例抑制与饱和度 if isfail wc_z = targetsz(2) + opts.contextAmount*sum(targetsz); hc_z = targetsz(1) + opts.contextAmount*sum(targetsz); s_z = sqrt(wc_z*hc_z); scale_z = opts.exemplarSize / s_z; d_search = (opts.instanceSize - opts.exemplarSize)/2; pad = d_search/scale_z; s_x = s_z + 2*pad; isfail = 0; else s_x = max(min_s_x, min(max_s_x, (1-opts.scaleLR)*s_x + opts.scaleLR*scaledInstance(newScale))); targetsz = (1-opts.scaleLR)*targetsz + opts.scaleLR*[scaledTarget(1,newScale) scaledTarget(2,newScale)]; end else isfail = 1; s_x = max(min_s_x, min(max_s_x, s_x*1.1)); net_conv.layers(net_conv.getLayerIndex('circonv1_1')).block.enable = false; net_conv.layers(net_conv.getLayerIndex('circonv1_2')).block.enable = false; if strcmp(opts.nettype,'1res') net_conv.layers(net_conv.getLayerIndex('circonv2_1')).block.enable = false; net_conv.layers(net_conv.getLayerIndex('circonv2_2')).block.enable = false; end end else % scale damping and saturation s_x = max(min_s_x, min(max_s_x, (1-opts.scaleLR)*s_x + opts.scaleLR*scaledInstance(newScale))); targetsz = (1-opts.scaleLR)*targetsz + opts.scaleLR*[scaledTarget(1,newScale) scaledTarget(2,newScale)]; end % validate tmp = pos+targetsz./2; if tmp(1)<0||tmp(2)<0||tmp(1)>opts.imgsz(1)||tmp(2)>opts.imgsz(2) state.obj.failframes = state.obj.failframes+1; if state.obj.failframes>=2 pos = [size(img,1),size(img,2)]./2; state.obj.failframes =0; end isfail = 1; net_conv.layers(net_conv.getLayerIndex('circonv1_1')).block.enable = false; net_conv.layers(net_conv.getLayerIndex('circonv1_2')).block.enable = false; if strcmp(opts.nettype,'1res') net_conv.layers(net_conv.getLayerIndex('circonv2_1')).block.enable = false; net_conv.layers(net_conv.getLayerIndex('circonv2_2')).block.enable = false; end end targetsz = targetsz./opts.targetszrate; state.obj.s_x = s_x; state.obj.pos = pos; state.obj.targetsz = targetsz; state.obj.targetLoc = [pos([2,1]) - targetsz([2,1])/2, targetsz([2,1])]; state.seq.time = state.seq.time + toc; state.seq.isfail = isfail; end function [change_alpahf,change_featf] = update_change(corrfeat,new_corrfeat,lambda,issum) if nargin<4 issum =false; end % leanring filter from corrfeat to new_corrfeat cos_window = hann(size(corrfeat,1)) * hann(size(corrfeat,2))'; tcorrfeat = bsxfun(@times, corrfeat, cos_window); corrfeatf = fft2(tcorrfeat); numcorr = numel(corrfeatf(:,:,1)); if ~issum kcorrfeatf = (corrfeatf .* conj(corrfeatf))./numcorr; else kcorrfeatf = sum(corrfeatf .* conj(corrfeatf),3)./numel(corrfeatf); end tnew_corrfeat = bsxfun(@times, new_corrfeat, cos_window); tnew_corrfeatf = fft2(tnew_corrfeat); alphaf = (tnew_corrfeatf./ (kcorrfeatf+ lambda)); m=0.9; change_alpahf =m*alphaf+(1-m)*tnew_corrfeat;%就改这两个就OK！！！！！！这一行是V change_featf = corrfeatf;%就改这两行就OK！！！！！这一行是W end % fast online learning for V %因为V和W求的是t-1帧的变化系数，所以从跟踪开始，在第三帧才进行计算第二帧的变化系数。 %因此在完成第二帧跟踪后，首先将exemplar（范例）的特征保存在变量corrfeat中，第二帧（t-1）帧的特征保存在tcorrfeat中，然后执行以下代码 function net = update_v(net,feats_1,feats_t,p) [alphaf,featf] = update_change(feats_1{1}(:,:,:,1),feats_t{1},p.v_lambda); net.params(net.getParamIndex('cir11_alphaf')).value = alphaf; net.params(net.getParamIndex('cir11_featf')).value = featf; net.layers(net.getLayerIndex('circonv1_1')).block.enable = true; if strcmp(p.nettype,'1res') [alphaf,featf] = update_change(feats_1{2}(:,:,:,1),feats_t{2},p.v1_lambda); net.params(net.getParamIndex('cir21_alphaf')).value = alphaf; net.params(net.getParamIndex('cir21_featf')).value = featf; net.layers(net.getLayerIndex('circonv2_1')).block.enable = true; end end % fast online learning for W %具体来说，在t-1帧跟踪后，我们有目标位置，可以将图像I（t-1）裁剪到以目标位置为中心的区域G(t-1)， %并使用相同大小的搜索区域Z(t-1)。然后我们将G(t-1)与高斯权重图相乘，得到G(横杠）(t-1)以正确地突出显示前景区域。我们需要学习W(t-1)，它鼓励G(t-1)的深层功能类似于G(横杠)(t-1)， function net = update_w(net,feats,p) [alphaf,featf] = update_change(feats{1}(:,:,:,1),feats{1}(:,:,:,2),p.w_lambda); net.params(net.getParamIndex('cir12_alphaf')).value = alphaf; net.params(net.getParamIndex('cir12_featf')).value = featf; net.layers(net.getLayerIndex('circonv1_2')).block.enable = true; if strcmp(p.nettype,'1res') [alphaf,featf] = update_change(feats{2}(:,:,:,1),feats{2}(:,:,:,2),p.w1_lambda); net.params(net.getParamIndex('cir22_alphaf')).value = alphaf; net.params(net.getParamIndex('cir22_featf')).value = featf; net.layers(net.getLayerIndex('circonv2_2')).block.enable = true; end end -