Object Recognition and Scene Understanding（七）OpenCV中Latent SVM模型檔案XML

從該blog開始，逐步介紹DPM + Latent SVM。關於OpenCV下DPM+Latent SVM簡單介紹參考上一篇博文：OpenCV Latent SVM Discriminatively Trained Part
Based Models for Object Detection

 

以Cat.xml（opencv安裝目錄下sample/c內）為例

<Model>

         <!-- Number of components -->

         <NumComponents>2</NumComponents>

         <!-- Number of features -->

         <P>31</P>

         <!-- Score threshold -->

         <ScoreThreshold>-1.0028649999999999</ScoreThreshold>

         <Component>

                   <!-- Root filterdescription -->

                   <RootFilter>

                            <!-- Dimensions-->

                            <sizeX>5</sizeX>

                            <sizeY>9</sizeY>

                            <!-- Weights(binary representation) -->

                            <Weights>

                                     ...

                            </Weights>

                            <!-- Linear termin score function -->

                            <LinearTerm>-2.2535784347835031</LinearTerm>

                   </RootFilter>

 

                   <!-- Part filtersdescription -->

                   <PartFilters>

                            <NumPartFilters>6</NumPartFilters>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                   </PartFilters>

         </Component>

         <Component>

                   <!-- Root filterdescription -->

                   <RootFilter>

                            <!-- Dimensions-->

                            <sizeX>5</sizeX>

                            <sizeY>9</sizeY>

                            <!-- Weights(binary representation) -->

                            <Weights>

                                     ...

                            </Weights>

 

                            <!-- Linear termin score function -->

                            <LinearTerm>-2.5835343890077622</LinearTerm>

                   </RootFilter>

                                     <!--Part filters description -->

                   <PartFilters>

                            <NumPartFilters>6</NumPartFilters>

                            <!-- Part filter? description -->

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>

                            <PartFilter>...</PartFilter>                                 

                   </PartFilters>

         </Component>

</Model>

 

PartFilter內部結構：

<PartFilter>

         <sizeX>6</sizeX>

         <sizeY>6</sizeY>

         <!-- Weights (binary representation)-->

         <Weights></Weights>

 

         <!-- Part filter offset -->

         <V>

                   <Vx>3</Vx>

                   <Vy>1</Vy>

         </V>

         <!-- Quadratic penalty functioncoefficients -->

         <Penalty>

                   <dx>0.0004031731821276</dx>

                   <dy>-0.0003745111759062</dy>

                   <dxx>0.0100010270581015</dxx>

                   <dyy>0.0205820897831230</dyy>

         </Penalty>

</PartFilter>

 

註：1.裡面的資料僅是為了說明，weight節點資料量太大，省略，PartFilter節點有點複雜，單列出來；

2.其中有些簡潔說明性文字，不是很全，PartFilter中Vx，Vy為HOG的位移。Penalty節點下的各項是為了計算deformation
model。