VOC-release4.01 DPM训练的model(mat)转为OpenCV latentsvm可以加载的model(xml)
2016-02-23 09:24
756 查看
摘自:http://www.xuebuyuan.com/601093.html
转换后的single component person model检测效果不错,而VOC2007 2009中的three-component person model的检测效果则非常差,可能是OpenCV对混合模型的加载不太支持。(我只关心person的检测,所以只测试了person的model转换后的效果)
function MAT2XMLmodel_401(matmodel, xmlfile)
% jelly 2013-08-12
% Convert *.mat format model in the source example in
% Discriminatively Trained Deformable Part Models "voc-release4.01"
% to opencv's latentSVM detect input format *.xml
if(nargin < 2)
matmodel = 'INRIA/inriaperson_final';
xmlfile = 'INRIA/inriaperson_final.xml';
end
load(matmodel);
fid = fopen(xmlfile, 'w');
fprintf(fid, '<Model>\n');
%%
%获取组件数
ncom = length(model.rules{model.start});
fprintf(fid, '\t<!-- Number of components -->\n');
fprintf(fid, '\t<NumComponents>%d</NumComponents>\n', ncom);
%获取特征维数,固定值31维。model中没有记录
nfeature = 31;
fprintf(fid, '\t<!-- Number of features -->\n');
fprintf(fid, '\t<P>%d</P>\n', nfeature);
%Score threshold=model.thresh
fprintf(fid, '\t<!-- Score threshold -->\n');
fprintf(fid, '\t<ScoreThreshold>%.16f</ScoreThreshold>\n', model.thresh);
layer = 1;
%对于每一个组件分别获取它的root filter、part filter、deformation filter
for icom = 1:ncom
fprintf(fid, '\t<Component>\n');
fprintf(fid, '\t\t<!-- Root filter description -->\n');
fprintf(fid, '\t\t<RootFilter>\n');
% attention: X,Y swap
rhs = model.rules{model.start}(icom).rhs;
% assume the root filter is first on the rhs of the start rules
if model.symbols(rhs(1)).type == 'T'
% handle case where there's no deformation model for the root
root = model.symbols(rhs(1)).filter;
else
% handle case where there is a deformation model for the root
root = model.symbols(model.rules{rhs(1)}(layer).rhs).filter;
end
filternum = root;
sizeX = model.filters(filternum).size(2);
sizeY = model.filters(filternum).size(1);
fprintf(fid, '\t\t\t<!-- Dimensions -->\n');
fprintf(fid, '\t\t\t<sizeX>%d</sizeX>\n', sizeX);
fprintf(fid, '\t\t\t<sizeY>%d</sizeY>\n', sizeY);
fprintf(fid, '\t\t\t<!-- Weights (binary representation) -->\n');
fprintf(fid, '\t\t\t<Weights>');
for iY = 1:sizeY
for iX = 1:sizeX
% original mat has 32 which is larger than nfeature=31 by 1
fwrite(fid, model.filters(filternum).w(iY,iX,1:nfeature), 'double'); % need verify
end
end
fprintf(fid, '\t\t\t</Weights>\n');
fprintf(fid, '\t\t\t<!-- Linear term in score function -->\n');
fprintf(fid, '\t\t\t<LinearTerm>%.16f</LinearTerm>\n', ... % need verify
model.rules{model.start}(icom).offset.w);
fprintf(fid, '\t\t</RootFilter>\n');
fprintf(fid, '\t\t<!-- Part filters description -->\n');
fprintf(fid, '\t\t<PartFilters>\n');
%在每个component内,获取part filter的个数,并获取每个part的参数
npart = length(model.rules{model.start}(icom).rhs) -1 ;
fprintf(fid, '\t\t\t<NumPartFilters>%d</NumPartFilters>\n', npart);
%%获取每个part的相关参数[dx,dy,ds]和penalty[dx dy dxx dyy]
for ipart = 2: npart+1
fprintf(fid, '\t\t\t<!-- Part filter ? description -->\n');
fprintf(fid, '\t\t\t<PartFilter>\n');
irule = model.rules{model.start}(icom).rhs(ipart);
filternum = model.symbols(model.rules{irule}.rhs).filter;
sizeX = model.filters(filternum).size(2);
sizeY = model.filters(filternum).size(1);
fprintf(fid, '\t\t\t\t<sizeX>%d</sizeX>\n', sizeX);
fprintf(fid, '\t\t\t\t<sizeY>%d</sizeY>\n', sizeY);
fprintf(fid, '\t\t\t\t<!-- Weights (binary representation) -->\n');
fprintf(fid, '\t\t\t\t<Weights>');
for iY = 1:sizeY
for iX = 1:sizeX
% original mat has 32 which is larger than nfeature=31 by 1
fwrite(fid, model.filters(filternum).w(iY,iX,1:nfeature), 'double'); % need verify
end
end
fprintf(fid, '\t\t\t\t</Weights>\n');
fprintf(fid, '\t\t\t\t<!-- Part filter offset -->\n');
fprintf(fid, '\t\t\t\t<V>\n');
fprintf(fid, '\t\t\t\t\t<Vx>%d</Vx>\n',model.rules{model.start}(icom).anchor{ipart}(1)+1); %[dx,dy,ds]
fprintf(fid, '\t\t\t\t\t<Vy>%d</Vy>\n',model.rules{model.start}(icom).anchor{ipart}(2)+1);
fprintf(fid, '\t\t\t\t</V>\n');
fprintf(fid, '\t\t\t\t<!-- Quadratic penalty function coefficients -->\n');
fprintf(fid, '\t\t\t\t<Penalty>\n');
fprintf(fid, '\t\t\t\t\t<dx>%.16f</dx>\n',model.rules{irule}.def.w(2));
fprintf(fid, '\t\t\t\t\t<dy>%.16f</dy>\n',model.rules{irule}.def.w(4));
fprintf(fid, '\t\t\t\t\t<dxx>%.16f</dxx>\n',model.rules{irule}.def.w(1));
fprintf(fid, '\t\t\t\t\t<dyy>%.16f</dyy>\n',model.rules{irule}.def.w(3));
fprintf(fid, '\t\t\t\t</Penalty>\n');
fprintf(fid, '\t\t\t</PartFilter>\n');
end
fprintf(fid, '\t\t</PartFilters>\n');
fprintf(fid, '\t</Component>\n');
end
fprintf(fid, '</Model>\n');
fclose(fid);
end
转换后的single component person model检测效果不错,而VOC2007 2009中的three-component person model的检测效果则非常差,可能是OpenCV对混合模型的加载不太支持。(我只关心person的检测,所以只测试了person的model转换后的效果)
function MAT2XMLmodel_401(matmodel, xmlfile)
% jelly 2013-08-12
% Convert *.mat format model in the source example in
% Discriminatively Trained Deformable Part Models "voc-release4.01"
% to opencv's latentSVM detect input format *.xml
if(nargin < 2)
matmodel = 'INRIA/inriaperson_final';
xmlfile = 'INRIA/inriaperson_final.xml';
end
load(matmodel);
fid = fopen(xmlfile, 'w');
fprintf(fid, '<Model>\n');
%%
%获取组件数
ncom = length(model.rules{model.start});
fprintf(fid, '\t<!-- Number of components -->\n');
fprintf(fid, '\t<NumComponents>%d</NumComponents>\n', ncom);
%获取特征维数,固定值31维。model中没有记录
nfeature = 31;
fprintf(fid, '\t<!-- Number of features -->\n');
fprintf(fid, '\t<P>%d</P>\n', nfeature);
%Score threshold=model.thresh
fprintf(fid, '\t<!-- Score threshold -->\n');
fprintf(fid, '\t<ScoreThreshold>%.16f</ScoreThreshold>\n', model.thresh);
layer = 1;
%对于每一个组件分别获取它的root filter、part filter、deformation filter
for icom = 1:ncom
fprintf(fid, '\t<Component>\n');
fprintf(fid, '\t\t<!-- Root filter description -->\n');
fprintf(fid, '\t\t<RootFilter>\n');
% attention: X,Y swap
rhs = model.rules{model.start}(icom).rhs;
% assume the root filter is first on the rhs of the start rules
if model.symbols(rhs(1)).type == 'T'
% handle case where there's no deformation model for the root
root = model.symbols(rhs(1)).filter;
else
% handle case where there is a deformation model for the root
root = model.symbols(model.rules{rhs(1)}(layer).rhs).filter;
end
filternum = root;
sizeX = model.filters(filternum).size(2);
sizeY = model.filters(filternum).size(1);
fprintf(fid, '\t\t\t<!-- Dimensions -->\n');
fprintf(fid, '\t\t\t<sizeX>%d</sizeX>\n', sizeX);
fprintf(fid, '\t\t\t<sizeY>%d</sizeY>\n', sizeY);
fprintf(fid, '\t\t\t<!-- Weights (binary representation) -->\n');
fprintf(fid, '\t\t\t<Weights>');
for iY = 1:sizeY
for iX = 1:sizeX
% original mat has 32 which is larger than nfeature=31 by 1
fwrite(fid, model.filters(filternum).w(iY,iX,1:nfeature), 'double'); % need verify
end
end
fprintf(fid, '\t\t\t</Weights>\n');
fprintf(fid, '\t\t\t<!-- Linear term in score function -->\n');
fprintf(fid, '\t\t\t<LinearTerm>%.16f</LinearTerm>\n', ... % need verify
model.rules{model.start}(icom).offset.w);
fprintf(fid, '\t\t</RootFilter>\n');
fprintf(fid, '\t\t<!-- Part filters description -->\n');
fprintf(fid, '\t\t<PartFilters>\n');
%在每个component内,获取part filter的个数,并获取每个part的参数
npart = length(model.rules{model.start}(icom).rhs) -1 ;
fprintf(fid, '\t\t\t<NumPartFilters>%d</NumPartFilters>\n', npart);
%%获取每个part的相关参数[dx,dy,ds]和penalty[dx dy dxx dyy]
for ipart = 2: npart+1
fprintf(fid, '\t\t\t<!-- Part filter ? description -->\n');
fprintf(fid, '\t\t\t<PartFilter>\n');
irule = model.rules{model.start}(icom).rhs(ipart);
filternum = model.symbols(model.rules{irule}.rhs).filter;
sizeX = model.filters(filternum).size(2);
sizeY = model.filters(filternum).size(1);
fprintf(fid, '\t\t\t\t<sizeX>%d</sizeX>\n', sizeX);
fprintf(fid, '\t\t\t\t<sizeY>%d</sizeY>\n', sizeY);
fprintf(fid, '\t\t\t\t<!-- Weights (binary representation) -->\n');
fprintf(fid, '\t\t\t\t<Weights>');
for iY = 1:sizeY
for iX = 1:sizeX
% original mat has 32 which is larger than nfeature=31 by 1
fwrite(fid, model.filters(filternum).w(iY,iX,1:nfeature), 'double'); % need verify
end
end
fprintf(fid, '\t\t\t\t</Weights>\n');
fprintf(fid, '\t\t\t\t<!-- Part filter offset -->\n');
fprintf(fid, '\t\t\t\t<V>\n');
fprintf(fid, '\t\t\t\t\t<Vx>%d</Vx>\n',model.rules{model.start}(icom).anchor{ipart}(1)+1); %[dx,dy,ds]
fprintf(fid, '\t\t\t\t\t<Vy>%d</Vy>\n',model.rules{model.start}(icom).anchor{ipart}(2)+1);
fprintf(fid, '\t\t\t\t</V>\n');
fprintf(fid, '\t\t\t\t<!-- Quadratic penalty function coefficients -->\n');
fprintf(fid, '\t\t\t\t<Penalty>\n');
fprintf(fid, '\t\t\t\t\t<dx>%.16f</dx>\n',model.rules{irule}.def.w(2));
fprintf(fid, '\t\t\t\t\t<dy>%.16f</dy>\n',model.rules{irule}.def.w(4));
fprintf(fid, '\t\t\t\t\t<dxx>%.16f</dxx>\n',model.rules{irule}.def.w(1));
fprintf(fid, '\t\t\t\t\t<dyy>%.16f</dyy>\n',model.rules{irule}.def.w(3));
fprintf(fid, '\t\t\t\t</Penalty>\n');
fprintf(fid, '\t\t\t</PartFilter>\n');
end
fprintf(fid, '\t\t</PartFilters>\n');
fprintf(fid, '\t</Component>\n');
end
fprintf(fid, '</Model>\n');
fclose(fid);
end
相关文章推荐
- python中使用OpenCV进行人脸检测的例子
- opencv 做人脸识别 opencv 人脸匹配分析
- 使用opencv拉伸图像扩大分辨率示例
- 基于C++实现kinect+opencv 获取深度及彩色数据
- OpenCV 2.4.3 C++ 平滑处理分析
- Python中使用OpenCV库来进行简单的气象学遥感影像计算
- 利用Python和OpenCV库将URL转换为OpenCV格式的方法
- python结合opencv实现人脸检测与跟踪
- Python环境搭建之OpenCV的步骤方法
- Python实现OpenCV的安装与使用示例
- 在树莓派2或树莓派B+上安装Python和OpenCV的教程
- opencv-python学习一--人脸检测
- 在Ubuntu上安装OpenCV3.0和Python-openCV的经历
- OpenCV配置,从来没有这么简单!
- ubuntu下opencv和qt的安装配置
- OpenCV学习笔记(二十五)——OpenCV图形界面设计Qt+VS2008
- 分享一些OpenCV实现立体视觉的经验
- 关于OpenCv图像变换与基本图形检测
- "应用程序正常初始化失败"-0xc0150002 解决办法
- DPM 2010 Error Code Catalog