libvigraimpex-dev 1.10.0+dfsg-11ubuntu2 / usr / include / vigra / multi_convolution.hxx

//-- -*- c++ -*-
/************************************************************************/
/*                                                                      */
/*               Copyright 2003 by Christian-Dennis Rahn                */
/*                        and Ullrich Koethe                            */
/*                                                                      */
/*    This file is part of the VIGRA computer vision library.           */
/*    The VIGRA Website is                                              */
/*        http://hci.iwr.uni-heidelberg.de/vigra/                       */
/*    Please direct questions, bug reports, and contributions to        */
/*        ullrich.koethe@iwr.uni-heidelberg.de    or                    */
/*        vigra@informatik.uni-hamburg.de                               */
/*                                                                      */
/*    Permission is hereby granted, free of charge, to any person       */
/*    obtaining a copy of this software and associated documentation    */
/*    files (the "Software"), to deal in the Software without           */
/*    restriction, including without limitation the rights to use,      */
/*    copy, modify, merge, publish, distribute, sublicense, and/or      */
/*    sell copies of the Software, and to permit persons to whom the    */
/*    Software is furnished to do so, subject to the following          */
/*    conditions:                                                       */
/*                                                                      */
/*    The above copyright notice and this permission notice shall be    */
/*    included in all copies or substantial portions of the             */
/*    Software.                                                         */
/*                                                                      */
/*    THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND    */
/*    EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES   */
/*    OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND          */
/*    NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT       */
/*    HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,      */
/*    WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING      */
/*    FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR     */
/*    OTHER DEALINGS IN THE SOFTWARE.                                   */                
/*                                                                      */
/************************************************************************/

#ifndef VIGRA_MULTI_CONVOLUTION_H
#define VIGRA_MULTI_CONVOLUTION_H

#include "separableconvolution.hxx"
#include "array_vector.hxx"
#include "multi_array.hxx"
#include "accessor.hxx"
#include "numerictraits.hxx"
#include "navigator.hxx"
#include "metaprogramming.hxx"
#include "multi_pointoperators.hxx"
#include "multi_math.hxx"
#include "functorexpression.hxx"
#include "tinyvector.hxx"
#include "algorithm.hxx"

namespace vigra
{

namespace detail
{

struct DoubleYielder
{
    const double value;
    DoubleYielder(double v, unsigned, const char *const) : value(v) {}
    DoubleYielder(double v)                              : value(v) {}
    void operator++() {}
    double operator*() const { return value; }
};

template <typename X>
struct IteratorDoubleYielder
{
    X it;
    IteratorDoubleYielder(X i, unsigned, const char *const) : it(i) {}
    IteratorDoubleYielder(X i)                              : it(i) {}
    void operator++() { ++it; }
    double operator*() const { return *it; }
};

template <typename X>
struct SequenceDoubleYielder
{
    typename X::const_iterator it;
    SequenceDoubleYielder(const X & seq, unsigned dim,
                          const char *const function_name = "SequenceDoubleYielder")
        : it(seq.begin())
    {
        if (seq.size() == dim)
            return;
        std::string msg = "(): Parameter number be equal to the number of spatial dimensions.";
        vigra_precondition(false, function_name + msg);
    }
    void operator++() { ++it; }
    double operator*() const { return *it; }
};

template <typename X>
struct WrapDoubleIterator
{
    typedef
    typename IfBool< IsConvertibleTo<X, double>::value,
        DoubleYielder,
        typename IfBool< IsIterator<X>::value || IsArray<X>::value,
            IteratorDoubleYielder<X>,
            SequenceDoubleYielder<X>
        >::type
    >::type type;
};

template <class Param1, class Param2, class Param3>
struct WrapDoubleIteratorTriple
{
    typename WrapDoubleIterator<Param1>::type sigma_eff_it;
    typename WrapDoubleIterator<Param2>::type sigma_d_it;
    typename WrapDoubleIterator<Param3>::type step_size_it;
    WrapDoubleIteratorTriple(Param1 sigma_eff, Param2 sigma_d, Param3 step_size)
        : sigma_eff_it(sigma_eff), sigma_d_it(sigma_d), step_size_it(step_size) {}
    void operator++()
    {
        ++sigma_eff_it;
        ++sigma_d_it;
        ++step_size_it;
    }
    double sigma_eff() const { return *sigma_eff_it; }
    double sigma_d() const { return *sigma_d_it; }
    double step_size() const { return *step_size_it; }
    static void sigma_precondition(double sigma, const char *const function_name)
    {
        if (sigma < 0.0)
        {
             std::string msg = "(): Scale must be positive.";
             vigra_precondition(false, function_name + msg);
        }
    }
    double sigma_scaled(const char *const function_name = "unknown function ") const
    {
        sigma_precondition(sigma_eff(), function_name);
        sigma_precondition(sigma_d(), function_name);
        double sigma_squared = sq(sigma_eff()) - sq(sigma_d());
        if (sigma_squared > 0.0)
        {
            return std::sqrt(sigma_squared) / step_size();
        }
        else
        {
             std::string msg = "(): Scale would be imaginary or zero.";
             vigra_precondition(false, function_name + msg);
             return 0;
        }
    }
};

template <unsigned dim>
struct multiArrayScaleParam
{
    typedef TinyVector<double, dim> p_vector;
    typedef typename p_vector::const_iterator return_type;
    p_vector vec;

    template <class Param>
    multiArrayScaleParam(Param val, const char *const function_name = "multiArrayScaleParam")
    {
        typename WrapDoubleIterator<Param>::type in(val, dim, function_name);
        for (unsigned i = 0; i != dim; ++i, ++in)
            vec[i] = *in;
    }
    return_type operator()() const
    {
        return vec.begin();
    }
    static void precondition(unsigned n_par, const char *const function_name = "multiArrayScaleParam")
    {
        char n[3] = "0.";
        n[0] += dim;
        std::string msg = "(): dimension parameter must be ";
        vigra_precondition(dim == n_par, function_name + msg + n);
    }
    multiArrayScaleParam(double v0, double v1, const char *const function_name = "multiArrayScaleParam")
    {
        precondition(2, function_name);
        vec = p_vector(v0, v1);
    }
    multiArrayScaleParam(double v0, double v1, double v2, const char *const function_name = "multiArrayScaleParam")
    {
        precondition(3, function_name);
        vec = p_vector(v0, v1, v2);
    }
    multiArrayScaleParam(double v0, double v1, double v2,  double v3, const char *const function_name = "multiArrayScaleParam")
    {
        precondition(4, function_name);
        vec = p_vector(v0, v1, v2, v3);
    }
    multiArrayScaleParam(double v0, double v1, double v2,  double v3, double v4, const char *const function_name = "multiArrayScaleParam")
    {
        precondition(5, function_name);
        vec = p_vector(v0, v1, v2, v3, v4);
    }
};

} // namespace detail

#define VIGRA_CONVOLUTION_OPTIONS(function_name, default_value, member_name) \
    template <class Param> \
    ConvolutionOptions & function_name(const Param & val) \
    { \
        member_name = ParamVec(val, "ConvolutionOptions::" #function_name); \
        return *this; \
    } \
    ConvolutionOptions & function_name() \
    { \
        member_name = ParamVec(default_value, "ConvolutionOptions::" #function_name); \
        return *this; \
    } \
    ConvolutionOptions & function_name(double v0, double v1) \
    { \
        member_name = ParamVec(v0, v1, "ConvolutionOptions::" #function_name); \
        return *this; \
    } \
    ConvolutionOptions & function_name(double v0, double v1, double v2) \
    { \
        member_name = ParamVec(v0, v1, v2, "ConvolutionOptions::" #function_name); \
        return *this; \
    } \
    ConvolutionOptions & function_name(double v0, double v1, double v2, double v3) \
    { \
        member_name = ParamVec(v0, v1, v2, v3, "ConvolutionOptions::" #function_name); \
        return *this; \
    } \
    ConvolutionOptions & function_name(double v0, double v1, double v2, double v3, double v4) \
    { \
        member_name = ParamVec(v0, v1, v2, v3, v4, "ConvolutionOptions::" #function_name); \
        return *this; \
    }


/** \brief  Options class template for convolutions.
 
  <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
  Namespace: vigra
  
  This class enables the calculation of scale space convolutions
  such as \ref gaussianGradientMultiArray() on data with anisotropic
  discretization. For these, the result of the ordinary calculation
  has to be multiplied by factors of \f$1/w^{n}\f$ for each dimension,
  where \f$w\f$ is the step size of the grid in said dimension and
  \f$n\f$ is the differential order of the convolution, e.g., 1 for
  gaussianGradientMultiArray(), and 0 for gaussianSmoothMultiArray(),
  respectively. Also for each dimension in turn, the convolution's scale
  parameter \f$\sigma\f$ has to be replaced by
  \f$\sqrt{\sigma_\mathrm{eff}^2 - \sigma_\mathrm{D}^2}\Big/w\f$,
  where \f$\sigma_\mathrm{eff}\f$ is the resulting effective filtering
  scale. The data is assumed to be already filtered by a 
  gaussian smoothing with the scale parameter \f$\sigma_\mathrm{D}\f$
  (such as by measuring equipment). All of the above changes are
  automatically employed by the convolution functions for <tt>MultiArray</tt>s
  if a corresponding options object is provided.

  The <tt>ConvolutionOptions</tt> class must be parameterized by the dimension
  <tt>dim</tt>
  of the <tt>MultiArray</tt>s on which it is used. The actual per-axis
  options are set by (overloaded) member functions explained below,
  or else default to neutral values corresponding to the absence of the
  particular option.
  
  All member functions set <tt>dim</tt> values of the respective convolution
  option, one for each dimension. They may be set explicitly by multiple
  arguments for up to five dimensions, or by a single argument to the same
  value for all dimensions. For the general case, a single argument that is
  either a C-syle array, an iterator, or a C++ standard library style
  sequence (such as <tt>std::vector</tt>, with member functions <tt>begin()</tt>
  and <tt>size()</tt>) supplies the option values for any number of dimensions.
  
  Note that the return value of all member functions is <tt>*this</tt>, which
  provides the mechanism for concatenating member function calls as shown below.

  <b>usage with explicit parameters:</b>

  \code
  ConvolutionOptions<2> opt = ConvolutionOptions<2>().stepSize(1, 2.3);
  \endcode
 
  <b>usage with arrays:</b>
 
  \code
  const double step_size[3] = { x_scale, y_scale, z_scale };
  ConvolutionOptions<3> opt = ConvolutionOptions<3>().stepSize(step_size);
  \endcode

  <b>usage with C++ standard library style sequences:</b>
 
  \code
  TinyVector<double, 4> step_size(1, 1, 2.0, 1.5);
  TinyVector<double, 4>  r_sigmas(1, 1, 2.3, 3.2);
  ConvolutionOptions<4> opt = ConvolutionOptions<4>().stepSize(step_size).resolutionStdDev(r_sigmas);
  \endcode

  <b>usage with iterators:</b>

  \code
  ArrayVector<double> step_size;
  step_size.push_back(0);
  step_size.push_back(3);
  step_size.push_back(4);
  ArrayVector<double>::iterator i = step_size.begin();
  ++i;
  ConvolutionOptions<2> opt = ConvolutionOptions<2>().stepSize(i);
  \endcode

  <b>general usage in a convolution function call:</b>

  \code
  MultiArray<3, double> test_image;
  MultiArray<3, double> out_image;
  
  double scale = 5.0;
  gaussianSmoothMultiArray(test_image, out_image, scale,
                           ConvolutionOptions<3>()
                              .stepSize        (1, 1, 3.2)
                              .resolutionStdDev(1, 1, 4)
                          );
  \endcode
 
*/
template <unsigned dim>
class ConvolutionOptions
{
  public:
    typedef typename MultiArrayShape<dim>::type Shape;
    typedef detail::multiArrayScaleParam<dim> ParamVec;
    typedef typename ParamVec::return_type    ParamIt;

    ParamVec sigma_eff;
    ParamVec sigma_d;
    ParamVec step_size;
    ParamVec outer_scale;
    double window_ratio;
    Shape from_point, to_point;
     
    ConvolutionOptions()
    : sigma_eff(0.0),
      sigma_d(0.0),
      step_size(1.0),
      outer_scale(0.0),
      window_ratio(0.0)
    {}

    typedef typename detail::WrapDoubleIteratorTriple<ParamIt, ParamIt, ParamIt>
        ScaleIterator;
    typedef typename detail::WrapDoubleIterator<ParamIt>::type
        StepIterator;

    ScaleIterator scaleParams() const
    {
        return ScaleIterator(sigma_eff(), sigma_d(), step_size());
    }
    StepIterator stepParams() const
    {
        return StepIterator(step_size());
    }

    ConvolutionOptions outerOptions() const
    {
        ConvolutionOptions outer = *this;
        // backward-compatible values:
        return outer.stdDev(outer_scale()).resolutionStdDev(0.0);
    }

    // Step size per axis.
    // Default: dim values of 1.0
    VIGRA_CONVOLUTION_OPTIONS(stepSize, 1.0, step_size)
#ifdef DOXYGEN
        /** Step size(s) per axis, i.e., the distance between two
            adjacent pixels. Required for <tt>MultiArray</tt>
            containing anisotropic data.
 
            Note that a convolution containing a derivative operator
            of order <tt>n</tt> results in a multiplication by 
            \f${\rm stepSize}^{-n}\f$ for each axis.
            Also, the above standard deviations
            are scaled according to the step size of each axis.
            Default value for the options object if this member function is not
            used: A value of 1.0 for each dimension.
        */
    ConvolutionOptions<dim> & stepSize(...);
#endif

    // Resolution standard deviation per axis.
    // Default: dim values of 0.0
    VIGRA_CONVOLUTION_OPTIONS(resolutionStdDev, 0.0, sigma_d)
#ifdef DOXYGEN
        /** Resolution standard deviation(s) per axis, i.e., a supposed
            pre-existing gaussian filtering by this value.
       
            The standard deviation actually used by the convolution operators
            is \f$\sqrt{{\rm sigma}^{2} - {\rm resolutionStdDev}^{2}}\f$ for each
            axis.
            Default value for the options object if this member function is not
            used: A value of 0.0 for each dimension.
        */
    ConvolutionOptions<dim> & resolutionStdDev(...);
#endif

    // Standard deviation of scale space operators.
    // Default: dim values of 0.0
    VIGRA_CONVOLUTION_OPTIONS(stdDev, 0.0, sigma_eff)
    VIGRA_CONVOLUTION_OPTIONS(innerScale, 0.0, sigma_eff)

#ifdef DOXYGEN
        /** Standard deviation(s) of scale space operators, or inner scale(s) for \ref structureTensorMultiArray().
        
            Usually not
            needed, since a single value for all axes may be specified as a parameter
            <tt>sigma</tt> to the call of
            an convolution operator such as \ref gaussianGradientMultiArray(), and
            anisotropic data requiring the use of the stepSize() member function.
            Default value for the options object if this member function is not
            used: A value of 0.0 for each dimension.
        */
    ConvolutionOptions<dim> & stdDev(...);

        /** Standard deviation(s) of scale space operators, or inner scale(s) for \ref structureTensorMultiArray().
        
            Usually not
            needed, since a single value for all axes may be specified as a parameter
            <tt>sigma</tt> to the call of
            an convolution operator such as \ref gaussianGradientMultiArray(), and
            anisotropic data requiring the use of the stepSize() member function.
            Default value for the options object if this member function is not
            used: A value of 0.0 for each dimension.
        */
    ConvolutionOptions<dim> & innerScale(...);
#endif

    // Outer scale, for structure tensor.
    // Default: dim values of 0.0
    VIGRA_CONVOLUTION_OPTIONS(outerScale, 0.0, outer_scale)
#ifdef DOXYGEN
        /** Standard deviation(s) of the second convolution of the
            structure tensor. 

            Usually not needed, since a single value for
            all axes may be specified as a parameter <tt>outerScale</tt> to
            the call of \ref structureTensorMultiArray(), and
            anisotropic data requiring the use of the stepSize() member
            function.
            Default value for the options object if this member function is not
            used: A value of 0.0 for each dimension.
        */
    ConvolutionOptions<dim> & outerScale(...);
#endif

        /** Size of the filter window as a multiple of the scale parameter. 

            This option is only used for Gaussian filters and their derivatives.
            By default, the window size of a Gaussian filter is automatically 
            determined such that the error resulting from restricting the 
            infinitely large Gaussian function to a finite size is minimized. 
            In particular, the window radius is determined as
            <tt>radius = round(3.0 * sigma + 0.5 * order)</tt>, where 'order' is the 
            desired derivative order. In some cases, it is desirable to trade off 
            accuracy for speed, and this function can be used to request a smaller
            window radius.
            
            Default: <tt>0.0</tt> (i.e. determine the window size automatically)
        */
    ConvolutionOptions<dim> & filterWindowSize(double ratio)
    {
        vigra_precondition(ratio >= 0.0,
            "ConvolutionOptions::filterWindowSize(): ratio must not be negative.");
        window_ratio = ratio;
        return *this;
    }

        /** Restrict the filter to a subregion of the input array. 

            This is useful for speeding up computations by ignoring irrelevant 
            areas in the array. <b>Note:</b> It is assumed that the output array
            of the convolution has the size given in this function.  Negative ROI 
            boundaries are interpreted relative to the end of the respective dimension 
            (i.e. <tt>if(to[k] < 0) to[k] += source.shape(k);</tt>).
            
            Default: <tt>from = Shape(), to = Shape()</tt> (i.e. use entire array)
        */
    ConvolutionOptions<dim> & subarray(Shape const & from, Shape const & to)
    {
        from_point = from;
        to_point = to;
        return *this;
    }
};

namespace detail
{

/********************************************************/
/*                                                      */
/*        internalSeparableConvolveMultiArray           */
/*                                                      */
/********************************************************/

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor, class KernelIterator>
void
internalSeparableConvolveMultiArrayTmp(
                      SrcIterator si, SrcShape const & shape, SrcAccessor src,
                      DestIterator di, DestAccessor dest, KernelIterator kit)
{
    enum { N = 1 + SrcIterator::level };

    typedef typename NumericTraits<typename DestAccessor::value_type>::RealPromote TmpType;
    typedef typename AccessorTraits<TmpType>::default_accessor TmpAcessor;

    // temporary array to hold the current line to enable in-place operation
    ArrayVector<TmpType> tmp( shape[0] );

    typedef MultiArrayNavigator<SrcIterator, N> SNavigator;
    typedef MultiArrayNavigator<DestIterator, N> DNavigator;
    
    TmpAcessor acc;

    {
        // only operate on first dimension here
        SNavigator snav( si, shape, 0 );
        DNavigator dnav( di, shape, 0 );
        
        for( ; snav.hasMore(); snav++, dnav++ )
        {
             // first copy source to tmp for maximum cache efficiency
             copyLine(snav.begin(), snav.end(), src, tmp.begin(), acc);

             convolveLine(srcIterRange(tmp.begin(), tmp.end(), acc),
                          destIter( dnav.begin(), dest ),
                          kernel1d( *kit ) );
        }
        ++kit;
    }

    // operate on further dimensions
    for( int d = 1; d < N; ++d, ++kit )
    {
        DNavigator dnav( di, shape, d );

        tmp.resize( shape[d] );

        for( ; dnav.hasMore(); dnav++ )
        {
             // first copy source to tmp since convolveLine() cannot work in-place
             copyLine(dnav.begin(), dnav.end(), dest, tmp.begin(), acc);

             convolveLine(srcIterRange(tmp.begin(), tmp.end(), acc),
                          destIter( dnav.begin(), dest ),
                          kernel1d( *kit ) );
        }
    }
}

/********************************************************/
/*                                                      */
/*         internalSeparableConvolveSubarray            */
/*                                                      */
/********************************************************/

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor, class KernelIterator>
void
internalSeparableConvolveSubarray(
                      SrcIterator si, SrcShape const & shape, SrcAccessor src,
                      DestIterator di, DestAccessor dest, KernelIterator kit,
                      SrcShape const & start, SrcShape const & stop)
{
    enum { N = 1 + SrcIterator::level };

    typedef typename NumericTraits<typename DestAccessor::value_type>::RealPromote TmpType;
    typedef MultiArray<N, TmpType> TmpArray;
    typedef typename TmpArray::traverser TmpIterator;
    typedef typename AccessorTraits<TmpType>::default_accessor TmpAcessor;
    
    SrcShape sstart, sstop, axisorder, tmpshape;
    TinyVector<double, N> overhead;
    for(int k=0; k<N; ++k)
    {
        sstart[k] = start[k] - kit[k].right();
        if(sstart[k] < 0)
            sstart[k] = 0;
        sstop[k] = stop[k] - kit[k].left();
        if(sstop[k] > shape[k])
            sstop[k] = shape[k];
        overhead[k] = double(sstop[k] - sstart[k]) / (stop[k] - start[k]);
    }
    
    indexSort(overhead.begin(), overhead.end(), axisorder.begin(), std::greater<double>());
    
    SrcShape dstart, dstop(sstop - sstart);
    dstop[axisorder[0]]  = stop[axisorder[0]] - start[axisorder[0]];
    
    // temporary array to hold the current line to enable in-place operation
    MultiArray<N, TmpType> tmp(dstop);

    typedef MultiArrayNavigator<SrcIterator, N> SNavigator;
    typedef MultiArrayNavigator<TmpIterator, N> TNavigator;
    
    TmpAcessor acc;

    {
        // only operate on first dimension here
        SNavigator snav( si, sstart, sstop, axisorder[0]);
        TNavigator tnav( tmp.traverser_begin(), dstart, dstop, axisorder[0]);
        
        ArrayVector<TmpType> tmpline(sstop[axisorder[0]] - sstart[axisorder[0]]);
        
        int lstart = start[axisorder[0]] - sstart[axisorder[0]];
        int lstop  = lstart + (stop[axisorder[0]] - start[axisorder[0]]);

        for( ; snav.hasMore(); snav++, tnav++ )
        {
            // first copy source to tmp for maximum cache efficiency
            copyLine(snav.begin(), snav.end(), src, tmpline.begin(), acc);
            
            convolveLine(srcIterRange(tmpline.begin(), tmpline.end(), acc),
                         destIter(tnav.begin(), acc),
                         kernel1d( kit[axisorder[0]] ), lstart, lstop);
        }
    }
    
    // operate on further dimensions
    for( int d = 1; d < N; ++d)
    {
        TNavigator tnav( tmp.traverser_begin(), dstart, dstop, axisorder[d]);
        
        ArrayVector<TmpType> tmpline(dstop[axisorder[d]] - dstart[axisorder[d]]);
        
        int lstart = start[axisorder[d]] - sstart[axisorder[d]];
        int lstop  = lstart + (stop[axisorder[d]] - start[axisorder[d]]);

        for( ; tnav.hasMore(); tnav++ )
        {
            // first copy source to tmp because convolveLine() cannot work in-place
            copyLine(tnav.begin(), tnav.end(), acc, tmpline.begin(), acc );

            convolveLine(srcIterRange(tmpline.begin(), tmpline.end(), acc),
                         destIter( tnav.begin() + lstart, acc ),
                         kernel1d( kit[axisorder[d]] ), lstart, lstop);
        }
        
        dstart[axisorder[d]] = lstart;
        dstop[axisorder[d]] = lstop;
    }
    
    copyMultiArray(tmp.traverser_begin()+dstart, stop-start, acc, di, dest);              
}


template <class K>
void 
scaleKernel(K & kernel, double a)
{
    for(int i = kernel.left(); i <= kernel.right(); ++i)
        kernel[i] = detail::RequiresExplicitCast<typename K::value_type>::cast(kernel[i] * a);
}


} // namespace detail

/** \addtogroup MultiArrayConvolutionFilters Convolution filters for multi-dimensional arrays.

    These functions realize a separable convolution on an arbitrary dimensional
    array that is specified by iterators (compatible to \ref MultiIteratorPage)
    and shape objects. It can therefore be applied to a wide range of data structures
    (\ref vigra::MultiArrayView, \ref vigra::MultiArray etc.).
*/
//@{

/********************************************************/
/*                                                      */
/*             separableConvolveMultiArray              */
/*                                                      */
/********************************************************/

/** \brief Separated convolution on multi-dimensional arrays.

    This function computes a separated convolution on all dimensions
    of the given multi-dimensional array. Both source and destination
    arrays are represented by iterators, shape objects and accessors.
    The destination array is required to already have the correct size.

    There are two variants of this functions: one takes a single kernel
    of type \ref vigra::Kernel1D which is then applied to all dimensions,
    whereas the other requires an iterator referencing a sequence of
    \ref vigra::Kernel1D objects, one for every dimension of the data.
    Then the first kernel in this sequence is applied to the innermost
    dimension (e.g. the x-axis of an image), while the last is applied to the
    outermost dimension (e.g. the z-axis in a 3D image).

    This function may work in-place, which means that <tt>source.data() == dest.data()</tt> is allowed.
    A full-sized internal array is only allocated if working on the destination
    array directly would cause round-off errors (i.e. if
    <tt>typeid(typename NumericTraits<T2>::RealPromote) != typeid(T2)</tt>).
    
    If <tt>start</tt> and <tt>stop</tt> have non-default values, they must represent
    a valid subarray of the input array. The convolution is then restricted to that 
    subarray, and it is assumed that the output array only refers to the
    subarray (i.e. <tt>dest.shape() == stop - start</tt>). Negative ROI boundaries are
    interpreted relative to the end of the respective dimension 
    (i.e. <tt>if(stop[k] < 0) stop[k] += source.shape(k);</tt>).

    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        // apply each kernel from the sequence 'kernels' in turn
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2, 
                  class KernelIterator>
        void
        separableConvolveMultiArray(MultiArrayView<N, T1, S1> const & source,
                                    MultiArrayView<N, T2, S2> dest, 
                                    KernelIterator kernels,
                                    typename MultiArrayShape<N>::type start = typename MultiArrayShape<N>::type(),
                                    typename MultiArrayShape<N>::type stop  = typename MultiArrayShape<N>::type());

        // apply the same kernel to all dimensions
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2, 
                  class T>
        void
        separableConvolveMultiArray(MultiArrayView<N, T1, S1> const & source,
                                    MultiArrayView<N, T2, S2> dest,
                                    Kernel1D<T> const & kernel,
                                    typename MultiArrayShape<N>::type const & start = typename MultiArrayShape<N>::type(),
                                    typename MultiArrayShape<N>::type const & stop = typename MultiArrayShape<N>::type());
    }
    \endcode

    \deprecatedAPI{separableConvolveMultiArray}
    pass \ref MultiIteratorPage "MultiIterators" and \ref DataAccessors :
    \code
    namespace vigra {
        // apply the same kernel to all dimensions
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor, class T>
        void
        separableConvolveMultiArray(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                    DestIterator diter, DestAccessor dest,
                                    Kernel1D<T> const & kernel,
                                    SrcShape const & start = SrcShape(),
                                    SrcShape const & stop = SrcShape());

        // apply each kernel from the sequence 'kernels' in turn
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor, class KernelIterator>
        void
        separableConvolveMultiArray(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                    DestIterator diter, DestAccessor dest,
                                    KernelIterator kernels,
                                    SrcShape const & start = SrcShape(),
                                    SrcShape const & stop = SrcShape());
    }
    \endcode
    use argument objects in conjunction with \ref ArgumentObjectFactories :
    \code
    namespace vigra {
        // apply the same kernel to all dimensions
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor, class T>
        void
        separableConvolveMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                    pair<DestIterator, DestAccessor> const & dest,
                                    Kernel1D<T> const & kernel,
                                    SrcShape const & start = SrcShape(),
                                    SrcShape const & stop = SrcShape());

        // apply each kernel from the sequence 'kernels' in turn
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor, class KernelIterator>
        void
        separableConvolveMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                    pair<DestIterator, DestAccessor> const & dest,
                                    KernelIterator kernels,
                                    SrcShape const & start = SrcShape(),
                                    SrcShape const & stop = SrcShape());
    }
    \endcode
    \deprecatedEnd

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, float>         dest(shape);
    ...
    Kernel1D<float> gauss;
    gauss.initGaussian(sigma);

    // smooth all dimensions with the same kernel
    separableConvolveMultiArray(source, dest, gauss);
    
    // create 3 Gauss kernels, one for each dimension, but smooth the z-axis less
    ArrayVector<Kernel1D<float> > kernels(3, gauss);
    kernels[2].initGaussian(sigma / 2.0);

    // perform Gaussian smoothing on all dimensions
    separableConvolveMultiArray(source, dest, kernels.begin());
    
    // create output array for a ROI
    MultiArray<3, float> destROI(shape - Shape3(10,10,10));
     
    // only smooth the given ROI (ignore 5 pixels on all sides of the array)
    separableConvolveMultiArray(source, destROI, gauss, Shape3(5,5,5), Shape3(-5,-5,-5));
    \endcode

    \deprecatedUsage{separableConvolveMultiArray}
    \code
    MultiArray<3, unsigned char>::size_type shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, float> dest(shape);
    ...
    Kernel1D<float> gauss;
    gauss.initGaussian(sigma);
    // create 3 Gauss kernels, one for each dimension
    ArrayVector<Kernel1D<float> > kernels(3, gauss);

    // perform Gaussian smoothing on all dimensions
    separableConvolveMultiArray(source, dest, 
                                kernels.begin());
    \endcode
    <b> Required Interface:</b>
    \code
    see \ref separableConvolveImage(), in addition:

    NumericTraits<T1>::RealPromote s = src[0];

    s = s + s;
    s = kernel(0) * s;
    \endcode
    \deprecatedEnd

    \see vigra::Kernel1D, convolveLine()
*/
doxygen_overloaded_function(template <...> void separableConvolveMultiArray)

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor, class KernelIterator>
void
separableConvolveMultiArray( SrcIterator s, SrcShape const & shape, SrcAccessor src,
                             DestIterator d, DestAccessor dest, 
                             KernelIterator kernels,
                             SrcShape start = SrcShape(),
                             SrcShape stop = SrcShape())
{
    typedef typename NumericTraits<typename DestAccessor::value_type>::RealPromote TmpType;


    if(stop != SrcShape())
    {
        
        enum { N = 1 + SrcIterator::level };
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(shape, start);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(shape, stop);
        
        for(int k=0; k<N; ++k)
            vigra_precondition(0 <= start[k] && start[k] < stop[k] && stop[k] <= shape[k],
              "separableConvolveMultiArray(): invalid subarray shape.");

        detail::internalSeparableConvolveSubarray(s, shape, src, d, dest, kernels, start, stop);
    }
    else if(!IsSameType<TmpType, typename DestAccessor::value_type>::boolResult)
    {
        // need a temporary array to avoid rounding errors
        MultiArray<SrcShape::static_size, TmpType> tmpArray(shape);
        detail::internalSeparableConvolveMultiArrayTmp( s, shape, src,
             tmpArray.traverser_begin(), typename AccessorTraits<TmpType>::default_accessor(), kernels );
        copyMultiArray(srcMultiArrayRange(tmpArray), destIter(d, dest));
    }
    else
    {
        // work directly on the destination array
        detail::internalSeparableConvolveMultiArrayTmp( s, shape, src, d, dest, kernels );
    }
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor, class T>
inline void
separableConvolveMultiArray( SrcIterator s, SrcShape const & shape, SrcAccessor src,
                             DestIterator d, DestAccessor dest,
                             Kernel1D<T> const & kernel,
                             SrcShape const & start = SrcShape(),
                             SrcShape const & stop = SrcShape())
{
    ArrayVector<Kernel1D<T> > kernels(shape.size(), kernel);

    separableConvolveMultiArray( s, shape, src, d, dest, kernels.begin(), start, stop);
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor, class KernelIterator>
inline void
separableConvolveMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                            pair<DestIterator, DestAccessor> const & dest, 
                            KernelIterator kit,
                            SrcShape const & start = SrcShape(),
                            SrcShape const & stop = SrcShape())
{
    separableConvolveMultiArray( source.first, source.second, source.third,
                                 dest.first, dest.second, kit, start, stop );
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor, class T>
inline void
separableConvolveMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                            pair<DestIterator, DestAccessor> const & dest,
                            Kernel1D<T> const & kernel,
                            SrcShape const & start = SrcShape(),
                            SrcShape const & stop = SrcShape())
{
    ArrayVector<Kernel1D<T> > kernels(source.second.size(), kernel);

    separableConvolveMultiArray( source.first, source.second, source.third,
                                 dest.first, dest.second, kernels.begin(), start, stop);
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2, 
          class KernelIterator>
inline void
separableConvolveMultiArray(MultiArrayView<N, T1, S1> const & source,
                            MultiArrayView<N, T2, S2> dest, 
                            KernelIterator kit,
                            typename MultiArrayShape<N>::type start = typename MultiArrayShape<N>::type(),
                            typename MultiArrayShape<N>::type stop = typename MultiArrayShape<N>::type())
{
    if(stop != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), start);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), stop);
        vigra_precondition(dest.shape() == (stop - start),
            "separableConvolveMultiArray(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(source.shape() == dest.shape(),
            "separableConvolveMultiArray(): shape mismatch between input and output.");
    }
    separableConvolveMultiArray( srcMultiArrayRange(source),
                                 destMultiArray(dest), kit, start, stop );
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2, 
          class T>
inline void
separableConvolveMultiArray(MultiArrayView<N, T1, S1> const & source,
                            MultiArrayView<N, T2, S2> dest,
                            Kernel1D<T> const & kernel,
                            typename MultiArrayShape<N>::type const & start = typename MultiArrayShape<N>::type(),
                            typename MultiArrayShape<N>::type const & stop = typename MultiArrayShape<N>::type())
{
    ArrayVector<Kernel1D<T> > kernels(N, kernel);
    separableConvolveMultiArray(source, dest, kernels.begin(), start, stop);
}

/********************************************************/
/*                                                      */
/*            convolveMultiArrayOneDimension            */
/*                                                      */
/********************************************************/

/** \brief Convolution along a single dimension of a multi-dimensional arrays.

    This function computes a convolution along one dimension (specified by
    the parameter <tt>dim</tt> of the given multi-dimensional array with the given
    <tt>kernel</tt>. The destination array must already have the correct size.

    If <tt>start</tt> and <tt>stop</tt> have non-default values, they must represent
    a valid subarray of the input array. The convolution is then restricted to that 
    subarray, and it is assumed that the output array only refers to the
    subarray (i.e. <tt>dest.shape() == stop - start</tt>). Negative ROI boundaries are
    interpreted relative to the end of the respective dimension 
    (i.e. <tt>if(stop[k] < 0) stop[k] += source.shape(k);</tt>).

    This function may work in-place, which means that <tt>source.data() == dest.data()</tt> is allowed.

    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2, 
                  class T>
        void
        convolveMultiArrayOneDimension(MultiArrayView<N, T1, S1> const & source,
                                       MultiArrayView<N, T2, S2> dest,
                                       unsigned int dim, 
                                       Kernel1D<T> const & kernel,
                                       typename MultiArrayShape<N>::type start = typename MultiArrayShape<N>::type(),
                                       typename MultiArrayShape<N>::type stop  = typename MultiArrayShape<N>::type());
    }
    \endcode

    \deprecatedAPI{convolveMultiArrayOneDimension}
    pass \ref MultiIteratorPage "MultiIterators" and \ref DataAccessors :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor, class T>
        void
        convolveMultiArrayOneDimension(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                       DestIterator diter, DestAccessor dest,
                                       unsigned int dim, vigra::Kernel1D<T> const & kernel,
                                       SrcShape const & start = SrcShape(),
                                       SrcShape const & stop = SrcShape());
    }
    \endcode
    use argument objects in conjunction with \ref ArgumentObjectFactories :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor, class T>
        void
        convolveMultiArrayOneDimension(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                       pair<DestIterator, DestAccessor> const & dest,
                                       unsigned int dim, vigra::Kernel1D<T> const & kernel,
                                       SrcShape const & start = SrcShape(),
                                       SrcShape const & stop = SrcShape());
    }
    \endcode
    \deprecatedEnd

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, float> dest(shape);
    ...
    Kernel1D<float> gauss;
    gauss.initGaussian(sigma);

    // perform Gaussian smoothing along dimension 1 (height)
    convolveMultiArrayOneDimension(source, dest, 1, gauss);
    \endcode

    \see separableConvolveMultiArray()
*/
doxygen_overloaded_function(template <...> void convolveMultiArrayOneDimension)

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor, class T>
void
convolveMultiArrayOneDimension(SrcIterator s, SrcShape const & shape, SrcAccessor src,
                               DestIterator d, DestAccessor dest,
                               unsigned int dim, vigra::Kernel1D<T> const & kernel,
                               SrcShape const & start = SrcShape(),
                               SrcShape const & stop = SrcShape())
{
    enum { N = 1 + SrcIterator::level };
    vigra_precondition( dim < N,
                        "convolveMultiArrayOneDimension(): The dimension number to convolve must be smaller "
                        "than the data dimensionality" );

    typedef typename NumericTraits<typename DestAccessor::value_type>::RealPromote TmpType;
    typedef typename AccessorTraits<TmpType>::default_const_accessor TmpAccessor;
    ArrayVector<TmpType> tmp( shape[dim] );

    typedef MultiArrayNavigator<SrcIterator, N> SNavigator;
    typedef MultiArrayNavigator<DestIterator, N> DNavigator;
    
    SrcShape sstart, sstop(shape), dstart, dstop(shape);
    
    if(stop != SrcShape())
    {
        sstart = start;
        sstop  = stop;
        sstart[dim] = 0;
        sstop[dim]  = shape[dim];
        dstop = stop - start;
    }

    SNavigator snav( s, sstart, sstop, dim );
    DNavigator dnav( d, dstart, dstop, dim );

    for( ; snav.hasMore(); snav++, dnav++ )
    {
        // first copy source to temp for maximum cache efficiency
        copyLine(snav.begin(), snav.end(), src,
                 tmp.begin(), typename AccessorTraits<TmpType>::default_accessor() );

        convolveLine(srcIterRange( tmp.begin(), tmp.end(), TmpAccessor()),
                     destIter( dnav.begin(), dest ),
                     kernel1d( kernel), start[dim], stop[dim]);
    }
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor, class T>
inline void
convolveMultiArrayOneDimension(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                               pair<DestIterator, DestAccessor> const & dest,
                               unsigned int dim,
                               Kernel1D<T> const & kernel,
                               SrcShape const & start = SrcShape(),
                               SrcShape const & stop = SrcShape())
{
    convolveMultiArrayOneDimension(source.first, source.second, source.third,
                                   dest.first, dest.second, dim, kernel, start, stop);
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2, 
          class T>
inline void
convolveMultiArrayOneDimension(MultiArrayView<N, T1, S1> const & source,
                               MultiArrayView<N, T2, S2> dest,
                               unsigned int dim, 
                               Kernel1D<T> const & kernel,
                               typename MultiArrayShape<N>::type start = typename MultiArrayShape<N>::type(),
                               typename MultiArrayShape<N>::type stop = typename MultiArrayShape<N>::type())
{
    if(stop != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), start);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), stop);
        vigra_precondition(dest.shape() == (stop - start),
            "convolveMultiArrayOneDimension(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(source.shape() == dest.shape(),
            "convolveMultiArrayOneDimension(): shape mismatch between input and output.");
    }
    convolveMultiArrayOneDimension(srcMultiArrayRange(source),
                                   destMultiArray(dest), dim, kernel, start, stop);
}

/********************************************************/
/*                                                      */
/*             gaussianSmoothMultiArray                 */
/*                                                      */
/********************************************************/

/** \brief Isotropic Gaussian smoothing of a multi-dimensional arrays.

    This function computes an isotropic convolution of the given N-dimensional
    array with a Gaussian filter at the given standard deviation <tt>sigma</tt>.
    Both source and destination arrays are represented by
    iterators, shape objects and accessors. The destination array is required to
    already have the correct size. This function may work in-place, which means
    that <tt>source.data() == dest.data()</tt> is allowed. It is implemented by a call to
    \ref separableConvolveMultiArray() with the appropriate kernel.

    Anisotropic data should be passed with appropriate
    \ref ConvolutionOptions, the parameter <tt>opt</tt> is otherwise optional
    unless the parameter <tt>sigma</tt> is omitted.

    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        // pass filter scale explicitly
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void
        gaussianSmoothMultiArray(MultiArrayView<N, T1, S1> const & source,
                                 MultiArrayView<N, T2, S2> dest,
                                 double sigma,
                                 ConvolutionOptions<N> opt = ConvolutionOptions<N>());

        // pass filer scale(s) in the option object
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void
        gaussianSmoothMultiArray(MultiArrayView<N, T1, S1> const & source,
                                 MultiArrayView<N, T2, S2> dest,
                                 ConvolutionOptions<N> opt);
    }
    \endcode

    \deprecatedAPI{gaussianSmoothMultiArray}
    pass \ref MultiIteratorPage "MultiIterators" and \ref DataAccessors :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        gaussianSmoothMultiArray(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                 DestIterator diter, DestAccessor dest,
                                 double sigma,
                                 const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    use argument objects in conjunction with \ref ArgumentObjectFactories :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        gaussianSmoothMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                 pair<DestIterator, DestAccessor> const & dest,
                                 double sigma,
                                 const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    \deprecatedEnd

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, float>         dest(shape);
    ...
    // perform isotropic Gaussian smoothing at scale 'sigma'
    gaussianSmoothMultiArray(source, dest, sigma);
    \endcode

    <b> Usage with anisotropic data:</b>

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, float>         dest(shape);
    TinyVector<float, 3> step_size;
    TinyVector<float, 3> resolution_sigmas;
    ...
    // perform anisotropic Gaussian smoothing at scale 'sigma'
    gaussianSmoothMultiArray(source, dest, sigma,
                             ConvolutionOptions<3>().stepSize(step_size).resolutionStdDev(resolution_sigmas));
    \endcode

    \see separableConvolveMultiArray()
*/
doxygen_overloaded_function(template <...> void gaussianSmoothMultiArray)

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
void
gaussianSmoothMultiArray( SrcIterator s, SrcShape const & shape, SrcAccessor src,
                   DestIterator d, DestAccessor dest,
                   const ConvolutionOptions<SrcShape::static_size> & opt,
                   const char *const function_name = "gaussianSmoothMultiArray" )
{
    static const int N = SrcShape::static_size;

    typename ConvolutionOptions<N>::ScaleIterator params = opt.scaleParams();
    ArrayVector<Kernel1D<double> > kernels(N);

    for (int dim = 0; dim < N; ++dim, ++params)
        kernels[dim].initGaussian(params.sigma_scaled(function_name), 1.0, opt.window_ratio);

    separableConvolveMultiArray(s, shape, src, d, dest, kernels.begin(), opt.from_point, opt.to_point);
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
gaussianSmoothMultiArray( SrcIterator s, SrcShape const & shape, SrcAccessor src,
                   DestIterator d, DestAccessor dest, double sigma,
                   const ConvolutionOptions<SrcShape::static_size> & opt = ConvolutionOptions<SrcShape::static_size>())
{
    ConvolutionOptions<SrcShape::static_size> par = opt;
    gaussianSmoothMultiArray(s, shape, src, d,  dest, par.stdDev(sigma));
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
gaussianSmoothMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                         pair<DestIterator, DestAccessor> const & dest,
                         const ConvolutionOptions<SrcShape::static_size> & opt)
{
    gaussianSmoothMultiArray( source.first, source.second, source.third,
                              dest.first, dest.second, opt );
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
gaussianSmoothMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                         pair<DestIterator, DestAccessor> const & dest, double sigma,
                         const ConvolutionOptions<SrcShape::static_size> & opt = ConvolutionOptions<SrcShape::static_size>())
{
    gaussianSmoothMultiArray( source.first, source.second, source.third,
                              dest.first, dest.second, sigma, opt );
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void
gaussianSmoothMultiArray(MultiArrayView<N, T1, S1> const & source,
                         MultiArrayView<N, T2, S2> dest,
                         ConvolutionOptions<N> opt)
{
    if(opt.to_point != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.from_point);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.to_point);
        vigra_precondition(dest.shape() == (opt.to_point - opt.from_point),
            "gaussianSmoothMultiArray(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(source.shape() == dest.shape(),
            "gaussianSmoothMultiArray(): shape mismatch between input and output.");
    }

    gaussianSmoothMultiArray( srcMultiArrayRange(source),
                              destMultiArray(dest), opt );
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void
gaussianSmoothMultiArray(MultiArrayView<N, T1, S1> const & source,
                         MultiArrayView<N, T2, S2> dest,
                         double sigma,
                         ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    gaussianSmoothMultiArray( source, dest, opt.stdDev(sigma) );
}


/********************************************************/
/*                                                      */
/*             gaussianGradientMultiArray               */
/*                                                      */
/********************************************************/

/** \brief Calculate Gaussian gradient of a multi-dimensional arrays.

    This function computes the Gaussian gradient of the given N-dimensional
    array with a sequence of first-derivative-of-Gaussian filters at the given
    standard deviation <tt>sigma</tt> (differentiation is applied to each dimension
    in turn, starting with the innermost dimension). The destination array is
    required to have a vector valued pixel type with as many elements as the number of
    dimensions. This function is implemented by calls to
    \ref separableConvolveMultiArray() with the appropriate kernels.

    Anisotropic data should be passed with appropriate
    \ref ConvolutionOptions, the parameter <tt>opt</tt> is otherwise optional
    unless the parameter <tt>sigma</tt> is omitted.

    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        // pass filter scale explicitly
        template <unsigned int N, class T1, class S1,
                  class T2, class S2>
        void
        gaussianGradientMultiArray(MultiArrayView<N, T1, S1> const & source,
                                   MultiArrayView<N, TinyVector<T2, N>, S2> dest,
                                   double sigma,
                                   ConvolutionOptions<N> opt = ConvolutionOptions<N>());

        // pass filter scale(s) in option object
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void
        gaussianGradientMultiArray(MultiArrayView<N, T1, S1> const & source,
                                   MultiArrayView<N, TinyVector<T2, N>, S2> dest,
                                   ConvolutionOptions<N> opt);
    }
    \endcode

    \deprecatedAPI{gaussianGradientMultiArray}
    pass \ref MultiIteratorPage "MultiIterators" and \ref DataAccessors :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        gaussianGradientMultiArray(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                   DestIterator diter, DestAccessor dest,
                                   double sigma, 
                                   const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    use argument objects in conjunction with \ref ArgumentObjectFactories :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        gaussianGradientMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                   pair<DestIterator, DestAccessor> const & dest,
                                   double sigma,
                                   const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    \deprecatedEnd

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, TinyVector<float, 3> > dest(shape);
    ...
    // compute Gaussian gradient at scale sigma
    gaussianGradientMultiArray(source, dest, sigma);
    \endcode

    <b> Usage with anisotropic data:</b>

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, TinyVector<float, 3> > dest(shape);
    TinyVector<float, 3> step_size;
    TinyVector<float, 3> resolution_sigmas;
    ...
    // compute Gaussian gradient at scale sigma
    gaussianGradientMultiArray(source, dest, sigma,
                               ConvolutionOptions<3>().stepSize(step_size).resolutionStdDev(resolution_sigmas));
    \endcode

    \see separableConvolveMultiArray()
*/
doxygen_overloaded_function(template <...> void gaussianGradientMultiArray)

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
void
gaussianGradientMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                           DestIterator di, DestAccessor dest,
                           ConvolutionOptions<SrcShape::static_size> const & opt,
                           const char *const function_name = "gaussianGradientMultiArray")
{
    typedef typename DestAccessor::value_type DestType;
    typedef typename DestType::value_type     DestValueType;
    typedef typename NumericTraits<DestValueType>::RealPromote KernelType;
   
    static const int N = SrcShape::static_size;
    typedef typename ConvolutionOptions<N>::ScaleIterator ParamType;

    for(int k=0; k<N; ++k)
        if(shape[k] <=0)
            return;

    vigra_precondition(N == (int)dest.size(di),
        "gaussianGradientMultiArray(): Wrong number of channels in output array.");

    ParamType params = opt.scaleParams();
    ParamType params2(params);

    ArrayVector<Kernel1D<KernelType> > plain_kernels(N);
    for (int dim = 0; dim < N; ++dim, ++params)
    {
        double sigma = params.sigma_scaled(function_name);
        plain_kernels[dim].initGaussian(sigma, 1.0, opt.window_ratio);
    }

    typedef VectorElementAccessor<DestAccessor> ElementAccessor;

    // compute gradient components
    for (int dim = 0; dim < N; ++dim, ++params2)
    {
        ArrayVector<Kernel1D<KernelType> > kernels(plain_kernels);
        kernels[dim].initGaussianDerivative(params2.sigma_scaled(), 1, 1.0, opt.window_ratio);
        detail::scaleKernel(kernels[dim], 1.0 / params2.step_size());
        separableConvolveMultiArray(si, shape, src, di, ElementAccessor(dim, dest), kernels.begin(), 
                                    opt.from_point, opt.to_point);
    }
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
void
gaussianGradientMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                           DestIterator di, DestAccessor dest, double sigma,
                           ConvolutionOptions<SrcShape::static_size> opt = ConvolutionOptions<SrcShape::static_size>())
{
    gaussianGradientMultiArray(si, shape, src, di, dest, opt.stdDev(sigma));
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
gaussianGradientMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                           pair<DestIterator, DestAccessor> const & dest,
                           ConvolutionOptions<SrcShape::static_size> const & opt )
{
    gaussianGradientMultiArray( source.first, source.second, source.third,
                                dest.first, dest.second, opt );
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
gaussianGradientMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                           pair<DestIterator, DestAccessor> const & dest,
                           double sigma,
                           const ConvolutionOptions<SrcShape::static_size> & opt = ConvolutionOptions<SrcShape::static_size>())
{
    gaussianGradientMultiArray( source.first, source.second, source.third,
                                dest.first, dest.second, sigma, opt );
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void
gaussianGradientMultiArray(MultiArrayView<N, T1, S1> const & source,
                           MultiArrayView<N, TinyVector<T2, N>, S2> dest,
                           ConvolutionOptions<N> opt )
{
    if(opt.to_point != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.from_point);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.to_point);
        vigra_precondition(dest.shape() == (opt.to_point - opt.from_point),
            "gaussianGradientMultiArray(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(source.shape() == dest.shape(),
            "gaussianGradientMultiArray(): shape mismatch between input and output.");
    }

    gaussianGradientMultiArray( srcMultiArrayRange(source),
                                destMultiArray(dest), opt );
}

template <unsigned int N, class T1, class S1,
          class T2, class S2>
inline void
gaussianGradientMultiArray(MultiArrayView<N, T1, S1> const & source,
                           MultiArrayView<N, TinyVector<T2, N>, S2> dest,
                           double sigma,
                           ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    gaussianGradientMultiArray( source, dest, opt.stdDev(sigma) );
}

/********************************************************/
/*                                                      */
/*              gaussianGradientMagnitude               */
/*                                                      */
/********************************************************/

namespace detail {

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
void 
gaussianGradientMagnitudeImpl(MultiArrayView<N+1, T1, S1> const & src,
                              MultiArrayView<N, T2, S2> dest,
                              ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    typename MultiArrayShape<N>::type shape(src.shape().template subarray<0,N>());
    if(opt.to_point != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(shape, opt.from_point);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(shape, opt.to_point);
        vigra_precondition(dest.shape() == (opt.to_point - opt.from_point),
            "gaussianGradientMagnitude(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(shape == dest.shape(),
            "gaussianGradientMagnitude(): shape mismatch between input and output.");
    }
              
    dest.init(0.0);
    
    typedef typename NumericTraits<T1>::RealPromote TmpType;
    MultiArray<N, TinyVector<TmpType, N> > grad(dest.shape());
    
    using namespace multi_math;
    
    for(int k=0; k<src.shape(N); ++k)
    {
        gaussianGradientMultiArray(src.bindOuter(k), grad, opt);
        
        dest += squaredNorm(grad);
    }
    dest = sqrt(dest);
}

} // namespace detail

    // documentation is in convolution.hxx
template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void 
gaussianGradientMagnitude(MultiArrayView<N+1, Multiband<T1>, S1> const & src,
                          MultiArrayView<N, T2, S2> dest,
                          ConvolutionOptions<N> const & opt)
{
    detail::gaussianGradientMagnitudeImpl<N, T1>(src, dest, opt);
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void 
gaussianGradientMagnitude(MultiArrayView<N, T1, S1> const & src,
                          MultiArrayView<N, T2, S2> dest,
                          ConvolutionOptions<N> const & opt)
{
    detail::gaussianGradientMagnitudeImpl<N, T1>(src.insertSingletonDimension(N), dest, opt);
}

template <unsigned int N, class T1, int M, class S1,
                          class T2, class S2>
inline void 
gaussianGradientMagnitude(MultiArrayView<N, TinyVector<T1, M>, S1> const & src,
                          MultiArrayView<N, T2, S2> dest,
                          ConvolutionOptions<N> const & opt)
{
    detail::gaussianGradientMagnitudeImpl<N, T1>(src.expandElements(N), dest, opt);
}

template <unsigned int N, class T1, unsigned int R, unsigned int G, unsigned int B, class S1,
                          class T2, class S2>
inline void 
gaussianGradientMagnitude(MultiArrayView<N, RGBValue<T1, R, G, B>, S1> const & src,
                          MultiArrayView<N, T2, S2> dest,
                          ConvolutionOptions<N> const & opt)
{
    detail::gaussianGradientMagnitudeImpl<N, T1>(src.expandElements(N), dest, opt);
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void 
gaussianGradientMagnitude(MultiArrayView<N, T1, S1> const & src,
                          MultiArrayView<N, T2, S2> dest,
                          double sigma,
                          ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    gaussianGradientMagnitude(src, dest, opt.stdDev(sigma));
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void 
gaussianGradientMagnitude(MultiArrayView<N+1, Multiband<T1>, S1> const & src,
                          MultiArrayView<N, T2, S2> dest,
                          double sigma,
                          ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    gaussianGradientMagnitude<N>(src, dest, opt.stdDev(sigma));
}

/********************************************************/
/*                                                      */
/*             symmetricGradientMultiArray              */
/*                                                      */
/********************************************************/

/** \brief Calculate gradient of a multi-dimensional arrays using symmetric difference filters.

    This function computes the gradient of the given N-dimensional
    array with a sequence of symmetric difference filters a (differentiation is applied
    to each dimension in turn, starting with the innermost dimension). 
    The destination array is required to have a vector valued pixel type with as many
    elements as the number of dimensions. This function is implemented by calls to
    \ref convolveMultiArrayOneDimension() with the symmetric difference kernel.

    Anisotropic data should be passed with appropriate
    \ref ConvolutionOptions, the parameter <tt>opt</tt> is optional
    otherwise.
    
    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void
        symmetricGradientMultiArray(MultiArrayView<N, T1, S1> const & source,
                                    MultiArrayView<N, TinyVector<T2, N>, S2> dest,
                                    ConvolutionOptions<N> opt = ConvolutionOptions<N>());
    }
    \endcode

    \deprecatedAPI{symmetricGradientMultiArray}
    pass \ref MultiIteratorPage "MultiIterators" and \ref DataAccessors :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        symmetricGradientMultiArray(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                    DestIterator diter, DestAccessor dest,
                                    const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    use argument objects in conjunction with \ref ArgumentObjectFactories :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        symmetricGradientMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                    pair<DestIterator, DestAccessor> const & dest,
                                    const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    \deprecatedEnd

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    MultiArray<3, unsigned char>::size_type shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, TinyVector<float, 3> > dest(shape);
    ...
    // compute gradient
    symmetricGradientMultiArray(srcMultiArrayRange(source), destMultiArray(dest));
    \endcode

    <b> Usage with anisotropic data:</b>

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, unsigned char> source(shape);
    MultiArray<3, TinyVector<float, 3> > dest(shape);
    TinyVector<float, 3> step_size;
    ...
    // compute gradient
    symmetricGradientMultiArray(source, dest,
                                ConvolutionOptions<3>().stepSize(step_size));
    \endcode

    \see convolveMultiArrayOneDimension()
*/
doxygen_overloaded_function(template <...> void symmetricGradientMultiArray)

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
void
symmetricGradientMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                            DestIterator di, DestAccessor dest,
                            const ConvolutionOptions<SrcShape::static_size> & opt = ConvolutionOptions<SrcShape::static_size>())
{
    typedef typename DestAccessor::value_type DestType;
    typedef typename DestType::value_type     DestValueType;
    typedef typename NumericTraits<DestValueType>::RealPromote KernelType;

    static const int N = SrcShape::static_size;
    typedef typename ConvolutionOptions<N>::StepIterator StepType;

    for(int k=0; k<N; ++k)
        if(shape[k] <=0)
            return;

    vigra_precondition(N == (int)dest.size(di),
        "symmetricGradientMultiArray(): Wrong number of channels in output array.");

    Kernel1D<KernelType> filter;
    filter.initSymmetricDifference();

    StepType step_size_it = opt.stepParams();

    typedef VectorElementAccessor<DestAccessor> ElementAccessor;

    // compute gradient components
    for (int d = 0; d < N; ++d, ++step_size_it)
    {
        Kernel1D<KernelType> symmetric(filter);
        detail::scaleKernel(symmetric, 1 / *step_size_it);
        convolveMultiArrayOneDimension(si, shape, src,
                                       di, ElementAccessor(d, dest),
                                       d, symmetric, opt.from_point, opt.to_point);
    }
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
symmetricGradientMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                            pair<DestIterator, DestAccessor> const & dest,
                            const ConvolutionOptions<SrcShape::static_size> & opt = ConvolutionOptions<SrcShape::static_size>())
{
    symmetricGradientMultiArray(source.first, source.second, source.third,
                                dest.first, dest.second, opt);
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void
symmetricGradientMultiArray(MultiArrayView<N, T1, S1> const & source,
                            MultiArrayView<N, TinyVector<T2, N>, S2> dest,
                            ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    if(opt.to_point != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.from_point);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.to_point);
        vigra_precondition(dest.shape() == (opt.to_point - opt.from_point),
            "symmetricGradientMultiArray(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(source.shape() == dest.shape(),
            "symmetricGradientMultiArray(): shape mismatch between input and output.");
    }

    symmetricGradientMultiArray(srcMultiArrayRange(source),
                                destMultiArray(dest), opt);
}

/********************************************************/
/*                                                      */
/*            laplacianOfGaussianMultiArray             */
/*                                                      */
/********************************************************/

/** \brief Calculate Laplacian of a N-dimensional arrays using Gaussian derivative filters.

    This function computes the Laplacian of the given N-dimensional
    array with a sequence of second-derivative-of-Gaussian filters at the given
    standard deviation <tt>sigma</tt>. Both source and destination 
    arrays must have scalar value_type. This function is implemented by calls to
    \ref separableConvolveMultiArray() with the appropriate kernels, followed by summation.

    Anisotropic data should be passed with appropriate
    \ref ConvolutionOptions, the parameter <tt>opt</tt> is otherwise optional
    unless the parameter <tt>sigma</tt> is left out.

    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        // pass scale explicitly
        template <unsigned int N, class T1, class S1,
                  class T2, class S2>
        void
        laplacianOfGaussianMultiArray(MultiArrayView<N, T1, S1> const & source,
                                      MultiArrayView<N, T2, S2> dest,
                                      double sigma,
                                      ConvolutionOptions<N> opt = ConvolutionOptions<N>());
        
        // pass scale(s) in option object
        template <unsigned int N, class T1, class S1,
                  class T2, class S2>
        void
        laplacianOfGaussianMultiArray(MultiArrayView<N, T1, S1> const & source,
                                      MultiArrayView<N, T2, S2> dest,
                                      ConvolutionOptions<N> opt );
    }
    \endcode

    \deprecatedAPI{laplacianOfGaussianMultiArray}
    pass \ref MultiIteratorPage "MultiIterators" and \ref DataAccessors :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        laplacianOfGaussianMultiArray(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                      DestIterator diter, DestAccessor dest,
                                      double sigma,
                                      const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    use argument objects in conjunction with \ref ArgumentObjectFactories :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        laplacianOfGaussianMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                      pair<DestIterator, DestAccessor> const & dest,
                                      double sigma,
                                      const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    \deprecatedEnd

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, float> source(shape);
    MultiArray<3, float> laplacian(shape);
    ...
    // compute Laplacian at scale sigma
    laplacianOfGaussianMultiArray(source, laplacian, sigma);
    \endcode

    <b> Usage with anisotropic data:</b>

    \code
    MultiArray<3, float> source(shape);
    MultiArray<3, float> laplacian(shape);
    TinyVector<float, 3> step_size;
    TinyVector<float, 3> resolution_sigmas;
    ...
    // compute Laplacian at scale sigma
    laplacianOfGaussianMultiArray(source, laplacian, sigma,
                                  ConvolutionOptions<3>().stepSize(step_size).resolutionStdDev(resolution_sigmas));
    \endcode

    \see separableConvolveMultiArray()
*/
doxygen_overloaded_function(template <...> void laplacianOfGaussianMultiArray)

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
void
laplacianOfGaussianMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                              DestIterator di, DestAccessor dest,
                              ConvolutionOptions<SrcShape::static_size> const & opt )
{ 
    using namespace functor;
    
    typedef typename DestAccessor::value_type DestType;
    typedef typename NumericTraits<DestType>::RealPromote KernelType;
    typedef typename AccessorTraits<KernelType>::default_accessor DerivativeAccessor;

    static const int N = SrcShape::static_size;
    typedef typename ConvolutionOptions<N>::ScaleIterator ParamType;
    
    ParamType params = opt.scaleParams();
    ParamType params2(params);

    ArrayVector<Kernel1D<KernelType> > plain_kernels(N);
    for (int dim = 0; dim < N; ++dim, ++params)
    {
        double sigma = params.sigma_scaled("laplacianOfGaussianMultiArray");
        plain_kernels[dim].initGaussian(sigma, 1.0, opt.window_ratio);
    }
    
    SrcShape dshape(shape);
    if(opt.to_point != SrcShape())
        dshape = opt.to_point - opt.from_point;
    
    MultiArray<N, KernelType> derivative(dshape);

    // compute 2nd derivatives and sum them up
    for (int dim = 0; dim < N; ++dim, ++params2)
    {
        ArrayVector<Kernel1D<KernelType> > kernels(plain_kernels);
        kernels[dim].initGaussianDerivative(params2.sigma_scaled(), 2, 1.0, opt.window_ratio);
        detail::scaleKernel(kernels[dim], 1.0 / sq(params2.step_size()));

        if (dim == 0)
        {
            separableConvolveMultiArray( si, shape, src, 
                                         di, dest, kernels.begin(), opt.from_point, opt.to_point);
        }
        else
        {
            separableConvolveMultiArray( si, shape, src, 
                                         derivative.traverser_begin(), DerivativeAccessor(), 
                                         kernels.begin(), opt.from_point, opt.to_point);
            combineTwoMultiArrays(di, dshape, dest, derivative.traverser_begin(), DerivativeAccessor(), 
                                  di, dest, Arg1() + Arg2() );
        }
    }
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
void
laplacianOfGaussianMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                              DestIterator di, DestAccessor dest, double sigma,
                              ConvolutionOptions<SrcShape::static_size> opt = ConvolutionOptions<SrcShape::static_size>())
{
    laplacianOfGaussianMultiArray(si, shape, src, di, dest, opt.stdDev(sigma));
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
laplacianOfGaussianMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                              pair<DestIterator, DestAccessor> const & dest,
                              ConvolutionOptions<SrcShape::static_size> const & opt )
{
    laplacianOfGaussianMultiArray( source.first, source.second, source.third,
                                   dest.first, dest.second, opt );
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
laplacianOfGaussianMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                              pair<DestIterator, DestAccessor> const & dest,
                              double sigma,
                              const ConvolutionOptions<SrcShape::static_size> & opt = ConvolutionOptions<SrcShape::static_size>())
{
    laplacianOfGaussianMultiArray( source.first, source.second, source.third,
                                   dest.first, dest.second,  sigma, opt );
}

template <unsigned int N, class T1, class S1,
          class T2, class S2>
inline void
laplacianOfGaussianMultiArray(MultiArrayView<N, T1, S1> const & source,
                              MultiArrayView<N, T2, S2> dest,
                              ConvolutionOptions<N> opt )
{
    if(opt.to_point != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.from_point);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.to_point);
        vigra_precondition(dest.shape() == (opt.to_point - opt.from_point),
            "laplacianOfGaussianMultiArray(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(source.shape() == dest.shape(),
            "laplacianOfGaussianMultiArray(): shape mismatch between input and output.");
    }

    laplacianOfGaussianMultiArray( srcMultiArrayRange(source),
                                   destMultiArray(dest), opt );
}

template <unsigned int N, class T1, class S1,
          class T2, class S2>
inline void
laplacianOfGaussianMultiArray(MultiArrayView<N, T1, S1> const & source,
                              MultiArrayView<N, T2, S2> dest,
                              double sigma,
                              ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    laplacianOfGaussianMultiArray( source, dest, opt.stdDev(sigma) );
}

/********************************************************/
/*                                                      */
/*             gaussianDivergenceMultiArray             */
/*                                                      */
/********************************************************/

/** \brief Calculate the divergence of a vector field using Gaussian derivative filters.

    This function computes the divergence of the given N-dimensional vector field
    with a sequence of first-derivative-of-Gaussian filters at the given
    standard deviation <tt>sigma</tt>. The input vector field can either be given as a sequence
    of scalar array views (one for each vector field component), represented by an
    iterator range, or by a single vector array with the appropriate shape.
    This function is implemented by calls to
    \ref separableConvolveMultiArray() with the suitable kernels, followed by summation.

    Anisotropic data should be passed with appropriate
    \ref ConvolutionOptions, the parameter <tt>opt</tt> is otherwise optional
    unless the parameter <tt>sigma</tt> is omitted.

    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        // specify input vector field as a sequence of scalar arrays
        template <class Iterator, 
                  unsigned int N, class T, class S>
        void 
        gaussianDivergenceMultiArray(Iterator vectorField, Iterator vectorFieldEnd,
                                     MultiArrayView<N, T, S> divergence,
                                     ConvolutionOptions<N> const & opt);
        
        template <class Iterator, 
                  unsigned int N, class T, class S>
        void 
        gaussianDivergenceMultiArray(Iterator vectorField, Iterator vectorFieldEnd,
                                     MultiArrayView<N, T, S> divergence,
                                     double sigma,
                                     ConvolutionOptions<N> opt = ConvolutionOptions<N>());
        
        // pass input vector field as an array of vectors
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void 
        gaussianDivergenceMultiArray(MultiArrayView<N, TinyVector<T1, N>, S1> const & vectorField,
                                     MultiArrayView<N, T2, S2> divergence,
                                     ConvolutionOptions<N> const & opt);
                                     
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void 
        gaussianDivergenceMultiArray(MultiArrayView<N, TinyVector<T1, N>, S1> const & vectorField,
                                     MultiArrayView<N, T2, S2> divergence,
                                     double sigma,
                                     ConvolutionOptions<N> opt = ConvolutionOptions<N>());
    }
    \endcode

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, TinyVector<float, 3> > source(shape);
    MultiArray<3, float> laplacian(shape);
    ...
    // compute divergence at scale sigma
    gaussianDivergenceMultiArray(source, laplacian, sigma);
    \endcode

    <b> Usage with anisotropic data:</b>

    \code
    MultiArray<3, TinyVector<float, 3> > source(shape);
    MultiArray<3, float> laplacian(shape);
    TinyVector<float, 3> step_size;
    TinyVector<float, 3> resolution_sigmas;
    ...
    // compute divergence at scale sigma
    gaussianDivergenceMultiArray(source, laplacian, sigma,
                                 ConvolutionOptions<3>().stepSize(step_size).resolutionStdDev(resolution_sigmas));
    \endcode
*/
doxygen_overloaded_function(template <...> void gaussianDivergenceMultiArray)

template <class Iterator, 
          unsigned int N, class T, class S>
void 
gaussianDivergenceMultiArray(Iterator vectorField, Iterator vectorFieldEnd,
                             MultiArrayView<N, T, S> divergence,
                             ConvolutionOptions<N> opt)
{
    typedef typename std::iterator_traits<Iterator>::value_type  ArrayType;
    typedef typename ArrayType::value_type                       SrcType;
    typedef typename NumericTraits<SrcType>::RealPromote         TmpType;
    typedef Kernel1D<double>                                     Kernel;
    
    vigra_precondition(std::distance(vectorField, vectorFieldEnd) == N,
        "gaussianDivergenceMultiArray(): wrong number of input arrays.");
    // more checks are performed in separableConvolveMultiArray()
    
    typename ConvolutionOptions<N>::ScaleIterator params = opt.scaleParams();
    ArrayVector<double> sigmas(N);
    ArrayVector<Kernel> kernels(N);
    for(unsigned int k = 0; k < N; ++k, ++params)
    {
        sigmas[k] = params.sigma_scaled("gaussianDivergenceMultiArray");
        kernels[k].initGaussian(sigmas[k], 1.0, opt.window_ratio);
    }
    
    MultiArray<N, TmpType> tmpDeriv(divergence.shape());
    
    for(unsigned int k=0; k < N; ++k, ++vectorField)
    {
        kernels[k].initGaussianDerivative(sigmas[k], 1, 1.0, opt.window_ratio);
        if(k == 0)
        {
            separableConvolveMultiArray(*vectorField, divergence, kernels.begin(), opt.from_point, opt.to_point);
        }
        else
        {
            separableConvolveMultiArray(*vectorField, tmpDeriv, kernels.begin(), opt.from_point, opt.to_point);
            divergence += tmpDeriv;
        }
        kernels[k].initGaussian(sigmas[k], 1.0, opt.window_ratio);
    }
}

template <class Iterator, 
          unsigned int N, class T, class S>
inline void 
gaussianDivergenceMultiArray(Iterator vectorField, Iterator vectorFieldEnd,
                             MultiArrayView<N, T, S> divergence,
                             double sigma,
                             ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    gaussianDivergenceMultiArray(vectorField, vectorFieldEnd, divergence, opt.stdDev(sigma));
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void 
gaussianDivergenceMultiArray(MultiArrayView<N, TinyVector<T1, N>, S1> const & vectorField,
                             MultiArrayView<N, T2, S2> divergence,
                             ConvolutionOptions<N> const & opt)
{
    ArrayVector<MultiArrayView<N, T1> > field;
    for(unsigned int k=0; k<N; ++k)
        field.push_back(vectorField.bindElementChannel(k));

    gaussianDivergenceMultiArray(field.begin(), field.end(), divergence, opt);
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void 
gaussianDivergenceMultiArray(MultiArrayView<N, TinyVector<T1, N>, S1> const & vectorField,
                             MultiArrayView<N, T2, S2> divergence,
                             double sigma,
                             ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    gaussianDivergenceMultiArray(vectorField, divergence, opt.stdDev(sigma));
}

/********************************************************/
/*                                                      */
/*              hessianOfGaussianMultiArray             */
/*                                                      */
/********************************************************/

/** \brief Calculate Hessian matrix of a N-dimensional arrays using Gaussian derivative filters.

    This function computes the Hessian matrix the given scalar N-dimensional
    array with a sequence of second-derivative-of-Gaussian filters at the given
    standard deviation <tt>sigma</tt>. The destination array must 
    have a vector valued element type with N*(N+1)/2 elements (it represents the
    upper triangular part of the symmetric Hessian matrix, flattened row-wise). 
    This function is implemented by calls to
    \ref separableConvolveMultiArray() with the appropriate kernels.

    Anisotropic data should be passed with appropriate
    \ref ConvolutionOptions, the parameter <tt>opt</tt> is otherwise optional
    unless the parameter <tt>sigma</tt> is omitted.

    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        // pass scale explicitly
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void
        hessianOfGaussianMultiArray(MultiArrayView<N, T1, S1> const & source,
                                    MultiArrayView<N, TinyVector<T2, int(N*(N+1)/2)>, S2> dest,
                                    double sigma,
                                    ConvolutionOptions<N> opt = ConvolutionOptions<N>());
        
        // pass scale(s) in option object
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void
        hessianOfGaussianMultiArray(MultiArrayView<N, T1, S1> const & source,
                                    MultiArrayView<N, TinyVector<T2, int(N*(N+1)/2)>, S2> dest,
                                    ConvolutionOptions<N> opt);
    }
    \endcode

    \deprecatedAPI{hessianOfGaussianMultiArray}
    pass \ref MultiIteratorPage "MultiIterators" and \ref DataAccessors :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        hessianOfGaussianMultiArray(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                    DestIterator diter, DestAccessor dest,
                                    double sigma,
                                    const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    use argument objects in conjunction with \ref ArgumentObjectFactories :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        hessianOfGaussianMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                    pair<DestIterator, DestAccessor> const & dest,
                                    double sigma,
                                    const ConvolutionOptions<N> & opt = ConvolutionOptions<N>());
    }
    \endcode
    \deprecatedEnd

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, float> source(shape);
    MultiArray<3, TinyVector<float, 6> > dest(shape);
    ...
    // compute Hessian at scale sigma
    hessianOfGaussianMultiArray(source, dest, sigma);
    \endcode

    <b> Usage with anisotropic data:</b>

    \code
    MultiArray<3, float> source(shape);
    MultiArray<3, TinyVector<float, 6> > dest(shape);
    TinyVector<float, 3> step_size;
    TinyVector<float, 3> resolution_sigmas;
    ...
    // compute Hessian at scale sigma
    hessianOfGaussianMultiArray(source, dest, sigma,
                                ConvolutionOptions<3>().stepSize(step_size).resolutionStdDev(resolution_sigmas));
    \endcode

    \see separableConvolveMultiArray(), vectorToTensorMultiArray()
*/
doxygen_overloaded_function(template <...> void hessianOfGaussianMultiArray)

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
void
hessianOfGaussianMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                            DestIterator di, DestAccessor dest,
                            ConvolutionOptions<SrcShape::static_size> const & opt )
{ 
    typedef typename DestAccessor::value_type DestType;
    typedef typename DestType::value_type     DestValueType;
    typedef typename NumericTraits<DestValueType>::RealPromote KernelType;

    static const int N = SrcShape::static_size;
    static const int M = N*(N+1)/2;
    typedef typename ConvolutionOptions<N>::ScaleIterator ParamType;
    
    for(int k=0; k<N; ++k)
        if(shape[k] <=0)
            return;

    vigra_precondition(M == (int)dest.size(di),
        "hessianOfGaussianMultiArray(): Wrong number of channels in output array.");

    ParamType params_init = opt.scaleParams();

    ArrayVector<Kernel1D<KernelType> > plain_kernels(N);
    ParamType params(params_init);
    for (int dim = 0; dim < N; ++dim, ++params)
    {
        double sigma = params.sigma_scaled("hessianOfGaussianMultiArray");
        plain_kernels[dim].initGaussian(sigma, 1.0, opt.window_ratio);
    }

    typedef VectorElementAccessor<DestAccessor> ElementAccessor;

    // compute elements of the Hessian matrix
    ParamType params_i(params_init);
    for (int b=0, i=0; i<N; ++i, ++params_i)
    {
        ParamType params_j(params_i);
        for (int j=i; j<N; ++j, ++b, ++params_j)
        {
            ArrayVector<Kernel1D<KernelType> > kernels(plain_kernels);
            if(i == j)
            {
                kernels[i].initGaussianDerivative(params_i.sigma_scaled(), 2, 1.0, opt.window_ratio);
            }
            else
            {
                kernels[i].initGaussianDerivative(params_i.sigma_scaled(), 1, 1.0, opt.window_ratio);
                kernels[j].initGaussianDerivative(params_j.sigma_scaled(), 1, 1.0, opt.window_ratio);
            }
            detail::scaleKernel(kernels[i], 1 / params_i.step_size());
            detail::scaleKernel(kernels[j], 1 / params_j.step_size());
            separableConvolveMultiArray(si, shape, src, di, ElementAccessor(b, dest),
                                        kernels.begin(), opt.from_point, opt.to_point);
        }
    }
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
hessianOfGaussianMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                            DestIterator di, DestAccessor dest, double sigma,
                            ConvolutionOptions<SrcShape::static_size> opt = ConvolutionOptions<SrcShape::static_size>())
{
    hessianOfGaussianMultiArray(si, shape, src, di, dest, opt.stdDev(sigma));
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
hessianOfGaussianMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                            pair<DestIterator, DestAccessor> const & dest,
                            ConvolutionOptions<SrcShape::static_size> const & opt )
{
    hessianOfGaussianMultiArray( source.first, source.second, source.third,
                                 dest.first, dest.second, opt );
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
hessianOfGaussianMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                            pair<DestIterator, DestAccessor> const & dest,
                            double sigma,
                            const ConvolutionOptions<SrcShape::static_size> & opt = ConvolutionOptions<SrcShape::static_size>())
{
    hessianOfGaussianMultiArray( source.first, source.second, source.third,
                                 dest.first, dest.second, sigma, opt );
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void
hessianOfGaussianMultiArray(MultiArrayView<N, T1, S1> const & source,
                            MultiArrayView<N, TinyVector<T2, int(N*(N+1)/2)>, S2> dest,
                            ConvolutionOptions<N> opt )
{
    if(opt.to_point != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.from_point);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.to_point);
        vigra_precondition(dest.shape() == (opt.to_point - opt.from_point),
            "hessianOfGaussianMultiArray(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(source.shape() == dest.shape(),
            "hessianOfGaussianMultiArray(): shape mismatch between input and output.");
    }

    hessianOfGaussianMultiArray( srcMultiArrayRange(source),
                                 destMultiArray(dest), opt );
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void
hessianOfGaussianMultiArray(MultiArrayView<N, T1, S1> const & source,
                            MultiArrayView<N, TinyVector<T2, int(N*(N+1)/2)>, S2> dest,
                            double sigma,
                            ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    hessianOfGaussianMultiArray( source, dest, opt.stdDev(sigma) );
}

namespace detail {

template<int N, class VectorType>
struct StructurTensorFunctor
{
    typedef VectorType result_type;
    typedef typename VectorType::value_type ValueType;
    
    template <class T>
    VectorType operator()(T const & in) const
    {
        VectorType res;
        for(int b=0, i=0; i<N; ++i)
        {
            for(int j=i; j<N; ++j, ++b)
            {
                res[b] = detail::RequiresExplicitCast<ValueType>::cast(in[i]*in[j]);
            }
        }
        return res;
    }
};

} // namespace detail

/********************************************************/
/*                                                      */
/*               structureTensorMultiArray              */
/*                                                      */
/********************************************************/

/** \brief Calculate th structure tensor of a multi-dimensional arrays.

    This function computes the gradient (outer product) tensor for each element
    of the given N-dimensional array with first-derivative-of-Gaussian filters at 
    the given <tt>innerScale</tt>, followed by Gaussian smoothing at <tt>outerScale</tt>.
    The destination array must have a vector valued pixel type with 
    N*(N+1)/2 elements (it represents the upper triangular part of the symmetric 
    structure tensor matrix, flattened row-wise). If the source array is also vector valued, the 
    resulting structure tensor is the sum of the individual tensors for each channel.
    This function is implemented by calls to
    \ref separableConvolveMultiArray() with the appropriate kernels.

    Anisotropic data should be passed with appropriate
    \ref ConvolutionOptions, the parameter <tt>opt</tt> is otherwise optional
    unless the parameters <tt>innerScale</tt> and <tt>outerScale</tt> are
    both omitted.

    <b> Declarations:</b>

    pass arbitrary-dimensional array views:
    \code
    namespace vigra {
        // pass scales explicitly
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void
        structureTensorMultiArray(MultiArrayView<N, T1, S1> const & source,
                                  MultiArrayView<N, TinyVector<T2, int(N*(N+1)/2)>, S2> dest,
                                  double innerScale, double outerScale,
                                  ConvolutionOptions<N> opt = ConvolutionOptions<N>());
        
        // pass scales in option object
        template <unsigned int N, class T1, class S1,
                                  class T2, class S2>
        void
        structureTensorMultiArray(MultiArrayView<N, T1, S1> const & source,
                                  MultiArrayView<N, TinyVector<T2, int(N*(N+1)/2)>, S2> dest, 
                                  ConvolutionOptions<N> opt );
    }
    \endcode

    \deprecatedAPI{structureTensorMultiArray}
    pass \ref MultiIteratorPage "MultiIterators" and \ref DataAccessors :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        structureTensorMultiArray(SrcIterator siter, SrcShape const & shape, SrcAccessor src,
                                  DestIterator diter, DestAccessor dest,
                                  double innerScale, double outerScale,
                                  ConvolutionOptions<N> opt);
    }
    \endcode
    use argument objects in conjunction with \ref ArgumentObjectFactories :
    \code
    namespace vigra {
        template <class SrcIterator, class SrcShape, class SrcAccessor,
                  class DestIterator, class DestAccessor>
        void
        structureTensorMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                                  pair<DestIterator, DestAccessor> const & dest,
                                  double innerScale, double outerScale,
                                  const ConvolutionOptions<N> & opt);
    }
    \endcode
    \deprecatedEnd

    <b> Usage:</b>

    <b>\#include</b> \<vigra/multi_convolution.hxx\><br/>
    Namespace: vigra

    \code
    Shape3 shape(width, height, depth);
    MultiArray<3, RGBValue<float> > source(shape);
    MultiArray<3, TinyVector<float, 6> > dest(shape);
    ...
    // compute structure tensor at scales innerScale and outerScale
    structureTensorMultiArray(source, dest, innerScale, outerScale);
    \endcode

    <b> Usage with anisotropic data:</b>

    \code
    MultiArray<3, RGBValue<float> > source(shape);
    MultiArray<3, TinyVector<float, 6> > dest(shape);
    TinyVector<float, 3> step_size;
    TinyVector<float, 3> resolution_sigmas;
    ...
    // compute structure tensor at scales innerScale and outerScale
    structureTensorMultiArray(source, dest, innerScale, outerScale,
                              ConvolutionOptions<3>().stepSize(step_size).resolutionStdDev(resolution_sigmas));
    \endcode

    \see separableConvolveMultiArray(), vectorToTensorMultiArray()
*/
doxygen_overloaded_function(template <...> void structureTensorMultiArray)

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
void
structureTensorMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                          DestIterator di, DestAccessor dest, 
                          ConvolutionOptions<SrcShape::static_size> opt)
{
    static const int N = SrcShape::static_size;
    static const int M = N*(N+1)/2;
    
    typedef typename DestAccessor::value_type DestType;
    typedef typename DestType::value_type     DestValueType;
    typedef typename NumericTraits<DestValueType>::RealPromote KernelType;
    typedef TinyVector<KernelType, N> GradientVector;
    typedef typename AccessorTraits<GradientVector>::default_accessor GradientAccessor;
    typedef typename AccessorTraits<DestType>::default_accessor GradientTensorAccessor;

    for(int k=0; k<N; ++k)
        if(shape[k] <=0)
            return;

    vigra_precondition(M == (int)dest.size(di),
        "structureTensorMultiArray(): Wrong number of channels in output array.");
        
    ConvolutionOptions<N> innerOptions = opt;
    ConvolutionOptions<N> outerOptions = opt.outerOptions();
    typename ConvolutionOptions<N>::ScaleIterator params = outerOptions.scaleParams();
    
    SrcShape gradientShape(shape);
    if(opt.to_point != SrcShape())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(shape, opt.from_point);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(shape, opt.to_point);
        
        for(int k=0; k<N; ++k, ++params)
        {
            Kernel1D<double> gauss;
            gauss.initGaussian(params.sigma_scaled("structureTensorMultiArray"), 1.0, opt.window_ratio);
            int dilation = gauss.right();
            innerOptions.from_point[k] = std::max<MultiArrayIndex>(0, opt.from_point[k] - dilation);
            innerOptions.to_point[k] = std::min<MultiArrayIndex>(shape[k], opt.to_point[k] + dilation);
        }
        outerOptions.from_point -= innerOptions.from_point;
        outerOptions.to_point -= innerOptions.from_point;
        gradientShape = innerOptions.to_point - innerOptions.from_point;
    }

    MultiArray<N, GradientVector> gradient(gradientShape);
    MultiArray<N, DestType> gradientTensor(gradientShape);
    gaussianGradientMultiArray(si, shape, src, 
                               gradient.traverser_begin(), GradientAccessor(), 
                               innerOptions,
                               "structureTensorMultiArray");

    transformMultiArray(gradient.traverser_begin(), gradientShape, GradientAccessor(), 
                        gradientTensor.traverser_begin(), GradientTensorAccessor(), 
                        detail::StructurTensorFunctor<N, DestType>());

    gaussianSmoothMultiArray(gradientTensor.traverser_begin(), gradientShape, GradientTensorAccessor(), 
                             di, dest, outerOptions,
                             "structureTensorMultiArray");
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
structureTensorMultiArray(SrcIterator si, SrcShape const & shape, SrcAccessor src,
                          DestIterator di, DestAccessor dest,
                          double innerScale, double outerScale,
                          ConvolutionOptions<SrcShape::static_size> opt = ConvolutionOptions<SrcShape::static_size>())
{
    structureTensorMultiArray(si, shape, src, di, dest,
                              opt.stdDev(innerScale).outerScale(outerScale));
}

template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
structureTensorMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                          pair<DestIterator, DestAccessor> const & dest, 
                          ConvolutionOptions<SrcShape::static_size> const & opt )
{
    structureTensorMultiArray( source.first, source.second, source.third,
                               dest.first, dest.second, opt );
}


template <class SrcIterator, class SrcShape, class SrcAccessor,
          class DestIterator, class DestAccessor>
inline void
structureTensorMultiArray(triple<SrcIterator, SrcShape, SrcAccessor> const & source,
                          pair<DestIterator, DestAccessor> const & dest,
                          double innerScale, double outerScale,
                          const ConvolutionOptions<SrcShape::static_size> & opt = ConvolutionOptions<SrcShape::static_size>())
{
    structureTensorMultiArray( source.first, source.second, source.third,
                               dest.first, dest.second,
                               innerScale, outerScale, opt);
}

template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void
structureTensorMultiArray(MultiArrayView<N, T1, S1> const & source,
                          MultiArrayView<N, TinyVector<T2, int(N*(N+1)/2)>, S2> dest, 
                          ConvolutionOptions<N> opt )
{
    if(opt.to_point != typename MultiArrayShape<N>::type())
    {
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.from_point);
        detail::RelativeToAbsoluteCoordinate<N-1>::exec(source.shape(), opt.to_point);
        vigra_precondition(dest.shape() == (opt.to_point - opt.from_point),
            "structureTensorMultiArray(): shape mismatch between ROI and output.");
    }
    else
    {
        vigra_precondition(source.shape() == dest.shape(),
            "structureTensorMultiArray(): shape mismatch between input and output.");
    }

    structureTensorMultiArray( srcMultiArrayRange(source),
                               destMultiArray(dest), opt );
}


template <unsigned int N, class T1, class S1,
                          class T2, class S2>
inline void
structureTensorMultiArray(MultiArrayView<N, T1, S1> const & source,
                          MultiArrayView<N, TinyVector<T2, int(N*(N+1)/2)>, S2> dest,
                          double innerScale, double outerScale,
                          ConvolutionOptions<N> opt = ConvolutionOptions<N>())
{
    structureTensorMultiArray(source, dest, opt.innerScale(innerScale).outerScale(outerScale));
}

//@}

} //-- namespace vigra


#endif        //-- VIGRA_MULTI_CONVOLUTION_H