Merge branch 'mpd'

MMVII/Doc/Programmer/Interpolators.tex

@@ -666,34 +666,99 @@ \subsubsection{{\tt GetValueAndGradInterpol}}
 
     % = = = = = = = = = = = = = = = = = = = = = = =
 
-\subsection{Interpolator and images optimization}
+\subsection{Usage of interpolators}
 
-%%        bool InsideInterpolator(const cInterpolator1D &,const cPtxd<double,Dim> & aP,tREAL8 aMargin=0.0) const; ///< Inside for Bilin
 
+We give a brieve description of possible usage of interpolators.
 
+   %   -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -
 
-%% GetVBl +ValInterpol
-    % = = = = = = = = = = = = = = = = = = = = = = =
+\subsubsection{Image ressampling}
+
+The classical usage of interpolators , nothing special to say, can be used for computing explicitely 
+an image transformation or "on the fly" for example in image correlation.
+
+   %   -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -
+
+\subsubsection{Interpolator and gradient}
+
+The method {\tt GetValueAndGradInterpol} can be used for computing the gradient in place
+of other classical algorithm (sobel, deriche \dots), don't know the value of it, by the way
+can give it a try.
 
+Perhaps more interesting is when we want to compute simultaneously a geometric transformation $\phi$
+and the gradient of the transformed image, if may be interesting for each pixel $p$ to :
 
+\begin{itemize}
+   \item computing simultaneously $I(\phi(p))$ and its gradient;
+   \item use the jacobian on $\phi(p)$ for infering the gradient in transformed image.
+\end{itemize}
 
+   %   -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -  -
 
-%% cDiffInterpolator1D * cDiffInterpolator1D::AllocFromNames(const std::vector<std::string> & aVName)
+\subsubsection{Interpolator and images optimization}
 
+When image are use in non linear optmisation (see~\ref{ImageOptDiff}), it is necessary to consider it like a
+continuous function $\RR^2 \rightarrow \RR$. In this case there is two option in
+\PPP (see \ref{SampleImageDiff}) :
 
-%% Image differ
+\begin{itemize}
+     \item  create an explicit interpolation function on formulas, see ~\ref{InterpBilCase};
+     \item  make a local linear approximation (see~\ref{ImDifGradMode}), in this case the
+            method {\tt GetValueAndGradInterpol} can be on option.
+\end{itemize}
 
     % = = = = = = = = = = = = = = = = = = = = = = =
     % = = = = = = = = = = = = = = = = = = = = = = =
     % = = = = = = = = = = = = = = = = = = = = = = =
 
 \section{Maintener's view}
 
+By maintener's we mean personn that would : make bug correction or make evolve the library .
+
 \label{InterpolMaintener}
 
-Efficiency 
-New interpol
-Bug correction
-Point partially inside
+    % = = = = = = = = = = = = = = = = = = = = = = =
+
+\subsection{Evolution}
+
+The following evolution can be envisaged :
+
+\begin{itemize}
+    \item \emph{add a new interpolator}, in this case it is recommanded to study the bench library describe
+	    bellow;
+    \item \emph{Efficiency} the main cost of computation are probably the internal loop  that compute
+	    the  scalar produtc between a line of  image and the coeffx in x (search {\tt SCALARPRODUCT}
+		in function {\tt GetValueInterpol, GetValueAndGradInterpol}
+
+    \item \emph{Efficiency} a minor enhancement would be to add a method that compute not a single
+	   weight, but, for a given phase $Ph$ compute the vector of weight for $Ph-k,Ph-k+1,\dots Ph+k$,
+           for tabulated interpolator some computation would be shared;
+
+   \item \emph{Partially inside} the fact that all the point must be inside the image with the sz of the kernel
+	   may be too restrictive, especially with high kernel, maybe write a method for points that are not
+		enough inside image.
+\end{itemize}
+
+    % = = = = = = = = = = = = = = = = = = = = = = =
+
+\subsection{Bench for interpolator}
+
+Several "bench" method have been written to check the correctness of interpolation. When making evolve the
+libray, they should be at least run again to check that there is no regression, and eventually completed,
+for example when adding new interpolator.
+
+\begin{itemize}
+     \item {\tt TplInterpol\_CmpLinearGetVBL}
+     \item {\tt TplInterpol\_CmpCubicFoncLinear}
+     \item {\tt TplInterpol\_FuncLowFreq}
+     \item {\tt BenchIntrinsiqOneInterpol}
+     \item {\tt BenchIntrinsiqMMVIIInterpol}
+\end{itemize}
+
+
+%% GetVBl +ValInterpol
+
+
 
 

MMVII/Doc/Programmer/NonLinearOptim.tex

@@ -770,6 +770,8 @@ \subsection{Compute next iteration}
 
 \section{Image, optmization and differenciation in MMVII}
 
+\label{ImageOptDiff}
+
           %  - - - - - - - - - - - - - - - - - - - 
 \subsection{Introduction}
 
@@ -817,6 +819,8 @@ \subsubsection{General formalization}
           %  - - - - - - - - - - - - - - - - - - - 
 \subsubsection{Bilinear case}
 
+\label{InterpBilCase}
+
 Let $p$ be a point and $q=(\tilde{x}_q,\tilde{y}_q)=\phi(p)$ its transformation by $\phi$. Let $x_0$ and $y_0$
 be the lower bound of $q$, and $x_1=x_0+1, y_1 = y_0+1$:
 
@@ -846,7 +850,7 @@ \subsubsection{Bicubic case}
           %  - - - - - - - - - - - - - - - - - - - 
 \subsubsection{Linear with gradient}
 
-\label{GradMode}
+\label{ImDifGradMode}
 
 Bi-linear and Bi-cubic case have the advantage to be a rigourous continuous analytical modelization of 
 image interpolated on whole $\RR^2$.  By the way bi-linear is known to be aliased, and bi-cubic
@@ -911,7 +915,7 @@ \subsubsection{For using generated code}
           %  - - - - - - - - - - - - - - - - - - - 
 \subsection{Facility functions for gradient  mode}
 
-For gradient mode corresponding to modelization of~\ref{GradMode}, we have very similar functions.
+For gradient mode corresponding to modelization of~\ref{ImDifGradMode}, we have very similar functions.
 For code generation we have :
 
 \begin{itemize}
@@ -933,6 +937,7 @@ \subsection{Facility functions for gradient  mode}
 %---------------------------------------------
 
 \section{Test example for image differenciation in MMVII}
+\label{SampleImageDiff}
 
 In this section we describe a test example, that has been implemented in MicMac, this example 
 has two objective : (1) make a didactic illustration of the fonctionnalities and (2) 
@@ -1048,10 +1053,11 @@ \subsection{Variant with gradient mode}
 make a list of adaptations  :
 
 \begin{itemize}
-   \item class {\tt cDeformImHomotethy} was added a member indicating if it works in mode gradient
+   \item in class {\tt cDeformImHomotethy} was added a member indicating if it has to works 
+	   in mode gradient;
    \item in the class {\tt cTestDeformIm},  there is a member indicating the interpolator if
-         it exist (else we ar in bilinear mode); searching all the occurence of {\tt mInterpol} in
-         this class will show the adaptation to do. 
+         it exist (else when  we are in bilinear mode); searching all the occurence of {\tt mInterpol} in
+         this class will show the difference between the two modes.
 \end{itemize}
 
 %---------------------------------------------

MMVII/src/Bench/BenchInterpolators.cpp

@@ -372,22 +372,24 @@ void  BenchInterpol(cParamExeBench & aParam)
 	delete aPtrTabInt2;
 
      }
+     // test that the two definition of cMMVII2Inperpol coincide (the spcialized with analyticall formula
+     // and the general, with paraméter 2, that solve a system (and test also with tab version)
      {
-         cMMVII2Inperpol aM2;
-         cMMVIIKInterpol aMK2(2.0);
-         cTabulatedInterpolator  aMKT2(cMMVIIKInterpol(2.0),1000,true);
+         cMMVII2Inperpol aM2;   //  version analytical formula
+         cMMVIIKInterpol aMK2(2.0);  // version solve system
+         cTabulatedInterpolator  aMKT2(cMMVIIKInterpol(2.0),1000,true); // tabulated
          for (tREAL8 aPh=-2.0 ; aPh<2.0 ; aPh+=0.1245)
          {
-            tREAL8 aD1 = std::abs( aM2.Weight(aPh)-aMK2.Weight(aPh));
-            tREAL8 aD2 = std::abs( aM2.Weight(aPh)-aMKT2.Weight(aPh));
+            tREAL8 aD1 = std::abs( aM2.Weight(aPh)-aMK2.Weight(aPh));  // Cmp analytical/solve syst
+            tREAL8 aD2 = std::abs( aM2.Weight(aPh)-aMKT2.Weight(aPh)); // Cmp analytical/tabulated
             // StdOut() << "M2PHHhh= " << aD1  << " " << aD2 << "\n";
 	    MMVII_INTERNAL_ASSERT_bench((aD1<1e-5) && (aD2<1e-5) ,"Interpol MMVII : Weigh=1");
          }
      }
 
      //  Bench on bicub interpola
      //
-     for (const auto & aP : {0.0,1.0,2.0,3.0})
+     for (const auto & aP : {0.0,1.0,2.0,3.0})  // Test different param of cubic
      {
          cCubicInterpolator  anI3(aP);
 

MMVII/src/ImagesBase/cIm2d_Interpolators.cpp

@@ -77,7 +77,7 @@ template <class Type>  tREAL8 cDataIm2D<Type>::GetValueInterpol(const cInterpola
 	// const tREAL8 *  aCurWX  = aLineWXInit;
 	tREAL8 aSomWIx = 0.0;
 
-	for (int aKX=0 ; aKX< aNbX ; aKX++)
+	for (int aKX=0 ; aKX< aNbX ; aKX++) // doc  SCALARPRODUCT
             aSomWIx += aLineIm[aKX]  * aLineWX[aKX] ;
 
 	/*  to see if this old style optim is any faster ? which I doubt ...
@@ -156,7 +156,7 @@ template <class Type>
 
 	tREAL8 aSomWIx = 0.0;
 	tREAL8 aSomDerWIx = 0.0;
-	for (int aKX=0 ; aKX< aNbX ; aKX++)
+	for (int aKX=0 ; aKX< aNbX ; aKX++) // doc  SCALARPRODUCT
 	{
             aSomWIx    += aLineIm[aKX]  *  aLineWX [aKX];
             aSomDerWIx += aLineIm[aKX]  *  aLineDerWX[aKX];