Generic Programming and Boost

1
Generic Programming and Boost

• Haifeng Gong

2
Contents

• Templates
• Generic Programming
• Boost C Library
• Boost CVPR
• Coding Advices

3
Templates
4
Templates

• Template class
• Template function
• Template class instantiation
• Template class specialization
• Template function overloading

5
Templates in C

• Great utility to programmers in C, especially
  when combined with multiple inheritance and
  operator overloading.
• The C Standard Template Library (STL) provides
  many useful functions within a framework of
  connected templates.
• Very expressive they may be used for things other
  than generic programming.
• Template metaprogramming, a way of pre-evaluating
  some of the code at compile-time rather than
  run-time

6
Template class - Specialization
convert_floatltTgttype convert_floatltchargttype c
onvert_floatltdoublegttype
template lttypename Tgt struct convert_float
typedef float type template ltgt struct
convert_floatltdoublegt typedef double
type template ltgt struct convert_floatltlonggt
typedef double type template ltgt struct
convert_floatltlong doublegt typedef long
double type
Specialization for double
Specialization for long
Specialization for long double
7
Template class Partial Specialization
CubicArrayltfloat, 3gt a CubicArrayltint, 2gt
b CubicArrayltstring, 2gt s
template lttypename T, size_t NumDimsgt struct
CubicArray CubicArrayltT, NumDims-1gt
elem16 CubicArrayltT, NumDims-1gt
operator(int i) return elemi
template lttypename Tgt struct CubicArrayltT,
1gt T elem16 T operator(int i)
return elemi
template ltgt struct CubicArrayltstring, 1gt
//...
template ltsize_t NumDimsgt struct
CubicArrayltstring, NumDimsgt CubicArrayltT,
NumDims-1gt elem16 CubicArrayltT, NumDims-1gt
operator(int i) return elemi void
clear_all()
8
Template function
templateltclass T, int Ngt T max(T t, N n) int
intArray -1,0,3,2,1 char charArray
'a', 'Z', 'n', 'c' int m max(intArray, 5)
//call template int max(int a, int b)
//Overloading int w max(5, 4)
//call max(a, b) char c
max(charArray, 4) //call
template templatelt templateltclassgt class T,
int Ngt Tltintgt max(Tltintgt t, N n)
9
Generic Programming
10
Generic Programming

• Introduction
• Concept
• Traits / Type Generators
• Policy Classes / Tag Dispatching
• Adaptors

11
Introduction

• Generic programming is a datatype-independent way
  of computer programming.
• The same source code will be used regardless of
  the datatype the code will be instantiated with
  or passed as parameters.

12
Generic programming

• Generic programming is about generalizing
  software components so that they can be easily
  reused in a wide variety of situations.
• In C, class and function templates are
  particularly effective mechanisms for generic
  programming because they make the generalization
  possible without sacrificing efficiency.

13
Concept

• Generic programming is "programming with
  concepts,"
• Concept is defined as a family of abstractions
  that are all related by a common set of
  requirements.

concept ForwardIteratorlttypename Itergt
typename value_type Iter operator(Iter
x) value_type operator(Iter) bool
operator(Iter, Iter) bool operator!(Iter,
Iter) //Not a valid syntax in C now
14
Policy / Tag

• Policy classes are small self-contained
  implementations for specific behaviors.
• Many of them are interchangeable which allows
  users to choose exactly which policies they wish.
  
• Tag is empty classes used as options for class or
  functions

15
Policy / Tag

• Policy example Allocator
• templateltclass _Ty, class _Ax allocatorlt_Tygt gt
• class list
• Tag example, fast compact, property tag
• struct fast_t //Tag
• struct compact_t //Tag
• void nearest_neighbor(..., fast_t opt)
• void nearest_neighbor(..., compact_t opt)

16
Traits

• Related types wrapper
• listltgtiterator,
• listltgtreference,
• listltgtvalue_type,
• unrefltgttype
• convert_floatltgttype
• graph_traitsltgtvertex_descriptor
• graph_traitsltgtedge_descriptor
• Hold relations between templated types

17
Adaptor

• Wrapper existed data structure to satisfy given
  concept
• For example, we can wrap CImage to satisfy
  boostMultiArray concept such as we can use
  boostMultiArray algorithms on CImage directly.

18
A real example
Template template parameter, the template type of
the input image
Template of the kernel
template lt templatelttypename, size_tgt class
MultiArray1, templatelttypename, size_tgt class
MultiArrayKer, templatelttypename, size_tgt class
MultiArray2, typename T1, typename TK, typename
T2 gt void filt( const MultiArray1ltT1, 2gt in,
const MultiArrayKerltTK, 2gt f,
MultiArray2ltT2, 2gt out, T1 (pad_pixel)(
const MultiArray1ltT1, 2gt, int,int) )
Template of the output image
Input image
kernel
Output image
Policy for how to padding pixels out of boundaries
19
Example contd
Type Traits
typedef typename convert_floatltT2gttype
TF2 int cy (f.shape()0-1)/2 int cx
(f.shape()1-1)/2 for(size_t yy0 yylt
in.shape()0 yy) for(size_t xx0
xxltin.shape()1 xx) TF2 t0
for(size_t dy0 dylt f.shape()0 dy)
for(size_t dx0 dxlt f.shape()1 dx)
t pad_pixel(in, yydy-cy,
xxdx-cx)fdydx
outyyxx static_castltT2gt(t)
MultiArray Concept has requirements
20
Boost
21
Boost

• Boost (http//www.boost.org) is a collection of
  free peer-reviewed portable C source libraries,
  which includes many useful algorithms and data
  structure.
• String and text processing/Input/Output
• Generic Programming/Template Metaprogramming
• Containers/Iterators/Algorithms
• Function Objects and higher-order programming
• Concurrent Programming
• Math and numerics
• Correctness and testing
• Data structures/Memory
• Parsing

22
Boost

• The aim of boost is to become part of C
  standard in the future. Ten Boost libraries are
  already included in the C Standards Committee's
  Library Technical Report (TR1) as a step toward
  becoming part of a future C Standard. More
  Boost libraries are proposed for the upcoming
  TR2.
• By taking advantages of the templates and generic
  programming, they are all very flexible and
  configurable. My favorite libraries include
  boostmulti_array, boostgraph , boostlambda,
  boostrandom.
• A good choice is to represent images and matrices
  in boostmulti_array, which will provide many
  useful features such as views and subarrays. So,
  if we want to reconstruct our codes, the first
  step is to modify the matrix algorithms to using
  boostmulti_array, and wrapper the image loader
  to produce a boostmulti_arrayltT,3gt image.

23
Boost

• Boost Multidimensional Array Library
• Boost Graph Library
• Boost Lamda Library

24
Boost multi_array Library

• Introduction
• Short Example
• Creating Views
• Facilities for CVPR

25
Boost multi_array Library

• The Boost MultiArray library provides a class
  template specifying an interface for
  N-dimensional containers.
• It includes a general array class template and
  array adaptors.
• The MultiArray class supports idiomatic array
  operations and interoperate with C Standard
  Library containers and algorithms.

26
Short Example
Instantiate the template
include ltboost/multi_array.hppgt using namespace
boost int main() typedef multi_arrayltfloat,3gt
 MultiArray typedef MultiArrayindex
 index  MultiArray   A(extents343) int v
alues    0 for(index  i    0 i  lt  3 i )
for(index  j  0 j  lt  4 j ) for(index 
k    0 k  lt  3 k ) Aijk   
values return 0
A signed integral type used for indexing into A
Create a 3D array that is 3 x 4 x 3
Access elements
27
Creating Views

• The sub-view supports the similar interface as
  the original array.
• It can retain the same number of dimensions as
  the original.
• It can have less dimensions than the original.

28
typedef  MultiArrayarray_viewlt3gt type 
view3D typedef  MultiArrayarray_viewlt2gt
type  view2D typedef  MultiArrayindex_range 
r   view3D aView3  A indicesr(0,2)r(0,4,
2)r()  view2D aView2  A indices0r()r()
     
The view type with Dims dimension template
array_viewltDimsgt type
This type specifies a range of indices over some
dimension of a MultiArray.
from beginning to end
base, stride, bound) 0,2,4) 0, 2
base,bound) 0,2) 0, 1
29
Facilities for CV

• The MultiArray container perfectly suits
  3D-graphics and digital image.
• For 3D-grahpics, slices are gained effortlessly
  with the tool of creating views.
• For digital images, we can easily choose one
  color channel similarly.

30
Facilities for CV

• By creating a view type of the same dimensions of
  the original, we can easily sample on a multigrid
  lattice.

e.g. use MultiArray to contain a color image
(256256) MultiArray   A(extents2562563
) create such a view view3D sample   A 
indicesr(0,256,4)r(0,256,4)r( ) 
31
Facilities for CV

• MultiArray provides a flexible process on the
  numbers of dimension.

e.g. (reshape in matlab) If we have a group of
images for training, Xi, with the size of m x
n. The MultiArray of 3 dimensions can be
created to contain this group. It is
convenient to train all these images (2D
(mn)xK), as well as to show the result of just
one image (3D mxnxK).
32
Boost Graph Library

• Generic Data Structures
• Generic Graph Algorithms

33
Generic Data Structures

• adjacency_list
• adjacency_matrix
• edge_list
• Suit directed, undirected, or bidirected graph
  when specially parameterized
• Provide containers for both node and edge

34
Generic Graph Algorithms

• Breadth First Search
• Depth First Search
• Uniform Cost Search
• Build blocks for constructing graph algorithms,
  currently including 13 in the BGL, such as
  connected components.

35
Boost Lambda Library

• Whats lambda
• Unnamed inplace function
• Lisp Programming Language
• If C has lambda

sort(v.begin(), v.end(), _1.timegt_2.time) transf
orm(v.begin(), v.end(), v2.begin(), sin(_1))
36
Boost Lambda Library

• Unnamed functions can be created easily with an
  intuitive syntax.
• e.g.
• for_each( a.begin(), a.end(), cout ltlt (1
  _1))
• Most of the restrictions in argument binding are
  removed, arbitrary arguments of practically any
  C function can be bound.
• Separate function composition operations are not
  needed, as function composition is supported
  implicitly.

_1 Placeholder for parameters
37
Lambda Examples

• For_each transform
• Sort by class member
• Pointers of each element

listltfloatgt v for_each(v.begin(), v.end(), _1
bind((float()(float))sin, _1))
struct Event //... int time vectorltEvent
gt v //... sort(v.begin(), v.end(),
bind(Eventtime, _1) gt bind(Eventtime, _2))
listltintgt v(10) vectorltintgt vp(10)
transform(v.begin(), v.end(), vp.begin(), _1)
38
Boost String

• Case-insensitive compare
• Lexical cast between string and number

equal(s1.begin(), s1.end(), s2.begin(),
is_iequal)
boostlexical_castltstdstringgt(2006) 2006
boostlexical_castltshortgt(2006) 2006
39
Boost CVPR
40
 Boost CVPR

• Boostcvpr uses boostmulti_arrayltgt to contain
  images and matrices, boostgraph ltgt to represent
  graph.
• The key feature that distinguished Boost CVPR
  from other related libraries like OpenCV is that
  it is a pure C library and takes full
  advantages of C templates and Boost generic
  programming.

41
multi_arrayltfloat,2gt gauKernel(extents55)
float sigma  1.5 gaussian(gauKernel,sigma)
coutltlt"gauKernel  "ltltendlltltgauKernelltlt""ltltendl
multi_arrayltfloat, 3gt in multi_arrayltfloat, 
3gt out load_image(in, "in.bmp")     
out.resize(extentsin.shape()0
in.shape()1in.shape()2)     
for(index cc  0 cc lt in.shape()2 cc)
    view2D filt_in  in indicesr()r()cc 
    view2D filt_out  out indicesr()r()cc
      filt(filt_in,gauKernel,filt_out)
save_image(out,"out.bmp")
Stream operator overloading
Load image
View a color channel
Filter a color channel
Save image
42
Color Space Transformation

• Function templates to transform RGB to LUV color
  space and in the reverse direction.

rgb2luv(rgbImageArray, luvImageArray) luv2rgb(luv
ImageArray, rgbImageArray)
43
Color Transformation

• This head file provides several components for
  color amendments.

nbin is the number of the color scales, and maxv
is the maximal value after process.

• Five useful functions
• saturate (in, lb, ub)
• histogram (in, hist, bin)
• histeq (in, out, nbin, maxv)
• level (in, out, bin, curv)
• autolevel (in, out, lb, ub)

p lt lb, let p lb p gt ub, let p ub.
curv bin are 1D multiarray of the same size,
guiding the leveling process.
hist bin are 1D multiarray of the same size.
histi store the statics data of the pixals
being the color of bini.
out can be omitted.
44
Array Visitor

• Wrapper of multi-folded loops for multiarray
• multi_for_each(_at)
• multi_copy(_at)
• multi_arith(_at)

eg.multi_arrayltfloat, 3gt  a(extents1001003)
multi_arrayltfloat, 3gt  b(extents1001003
) multi_arrayltfloat, 3gt  c(extents100100
3) multi_for_each(a, _12) //a 2
multi_copy(a, b, _22_1) //b 2a
multi_arith(a, b, c, _32_1_2)//c 2ba
boostmulti_arrayboostlambdaMatlabC
45
IO Stream

• By overloading the operator ltlt, the objects of
  multi_array, array_view, and sub_array types can
  be output at one time.

• Using the sentences just like (both file and std.
  console)
• coutltlt A cingtgt A
• foutltlt A fingtgt A

46
Advices on Coding
47
Code Separation

• There are 5 parts of algorithms and data
  structures in our program,
• (1) interface libraries such as MFC, dotNet, Qt,
  VCL, Xtreme, wxWindows
• (2) interface code of our own code
• (3) core algorithms and data structure oriented
  to our specified task
• (4) special CV and math algorithm libraries such
  as OpenCV, VisSDK, ImageMagick, gsl
• (5) the standard library
• We should make our core part as independent as
  possible. So there are 3 categories of calls or
  reference s between these parts, allowed,
  forbidden, and limitedly permitted.

48
Call Diagram
49
Code Separation

• The only directed call from the core is call the
  standard library.
• We often need to use some other CV algorithms,
  these algorithms should wrapper under our data
  structure before using, and we call the wrapper
  layer instead of the original library.
• Our core algorithms and data structures must be
  designed without the assumption that the
  interface be MFC, image loader be FreeImage,
  Canny algorithm be from OpenCV.

50
Forbidden Calls and References

• To make the code more separable, the following
  calls and references are forbidden
• (1) the core calls the interface or interface
  library algorithms directly, such as updating the
  progress bar directly during the iterations in
  the core
• (2) core refers the interface or interface
  library data structure, such as accessing
  MainFrame from the core directly, or using CRect
  and CString in the core data structure
  definition, or letting a core class inherited
  from MFC class.

51
Forbidden Calls and References
52
Callback Example
In MainFrame.h (VC/MFC)

• An example on iteration and progress

Or in MainFormUnit.h (CBuilder/VCL)
53
More Separation

• To make our work more reusable, another
  separation in the core is encouraged. We often
  implement generic algorithms in our projects,
  such as connected components, histogram
  equalization these codes should be independent
  from other task specified algorithms.

54
End

• Thanks
• Welcome to visit
• Video/Motion Group Reference Documents section _at_
  our SharePoint
• Some Materials are from Yiyi Weis seminar report

55
Research Platform

• Separation of Annotation Software and Research
  Code

56
Issues in ImageParser

• class static namespace
• Its better to define an object of class, just
  like App and MainFrame in MFC, though they just
  have one object, they do not use namespace
• Keep header files in order, using forward
  declaration
• If too many similar codes, abstraction is needed
  to make it neat and revisable

57
Interfacing

• 4 level of interfacing
• File exchange text, image, xml
• Executable file call
• Dll call
• Source code in one Executable file