Unconstrained Automatic Image Matching Using Multiresolutional CriticalPoint Filters Yoshihisa Shina - PowerPoint PPT Presentation

1 / 15
About This Presentation
Title:

Unconstrained Automatic Image Matching Using Multiresolutional CriticalPoint Filters Yoshihisa Shina

Description:

Unconstrained Automatic Image Matching Using Multiresolutional Critical-Point Filters ... The system tries to reduce the total energy by decreasing regardless of. ... – PowerPoint PPT presentation

Number of Views:28
Avg rating:3.0/5.0
Slides: 16
Provided by: visio5
Category:

less

Transcript and Presenter's Notes

Title: Unconstrained Automatic Image Matching Using Multiresolutional CriticalPoint Filters Yoshihisa Shina


1
Unconstrained Automatic Image Matching Using
Multiresolutional Critical-Point
FiltersYoshihisa Shinagawa, Tosiyasu L. Kunii
  • Jo, Young-Gwan
  • Computer Vision Lab.
  • CSE. POSTECH

2
Introduction
  • Image Matching
  • point, line, high-level feature matching
  • generating other views, morphing, volume
    rendering, 3D reconstruction, tracking
  • Hierarchical Matching
  • multiresolution
  • global matching

3
Notation
  • Level m the size of each image at the mth level
    is 2m?2m.
  • a pixel at the location (i,j) of the
    image at mth level. s0,1,2,3 indicating the
    kind of the critical-point filter.
  • intensity value of the pixel
  • submapping
    between p(m,s) and q(m,s)
  • is the parent of .
    is the child of .
  • ?,? weighting parameters, ?, ? penalty for
    violating BC

4
Critical-Point Filter
  • Maximum point
  • Minimum point
  • Saddle points

5
Overview of Image MatchingUsing
Multiresolutional Critical-Point Filters
  • Making subimages
  • critical-point filters
  • Selecting candidates for mapping ( or matching )
  • inherited quadrilateral
  • checking BC
  • if no pixel that satisfies BC exists
  • expand quadrilateral
  • abandon the third of BC
  • abandon the first and second of BC
  • Determining mapping
  • computing candidate minimizing energy ( or cost )

6
Bijectivity Conditions
  • The edges of the quadrilateral f (m,s)(R) should
    not intersect one another.
  • The orientation of the edges of should be the
    same as that of R ( clockwise or counterclockwise
    ).
  • The length of one edge of f (m,s)(R) can be zero
    to allow mappings that are retractions i.e., f
    (m,s)(R) may be a triangle. It is not allowed,
    however, to be a point nor a line segment (
    figures of area 0 ).

7
The Energy of the Mapping
  • Difference in the intensity of matched points
  • Difference in the location of matched points

8
Cost Related to the Pixel Intensity
  • The energy
  • The total energy

9
Cost Related to the Locationsof the Pixel for
Smooth Mapping
  • The energy
  • The total energy

10
Total Energy of the Mapping
  • The goal is to find a mapping that gives the
    minimum energy

11
Determining the Mappingwith Multiresolution
  • The value f (m,s)(i,j) (k,l) is determined
    using f (m-1,s) ( m1,2,,n ).
  • should be inside the inherited
    quadrilateral.
  • Inherited quadrilateral

12
The New Energy E0
  • The energy of submapping f(m,0)
  • The energy of submapping f(m,s) (s1,2,3)

13
Automatic Determinationof the Optimal Parameter
Values
  • Dynamic determination of ?
  • When we increase ?, a change in the mapping is
    not possible until becomes greater
    than one.
  • If exceeds one, some pixels may move
    to a stabler state.
  • When ? tries to go beyond the optimal value,
    however, the mapping becomes excessively
    distorted, and increases rapidly. In
    such a case, the total energy
    is dominated by . The system tries to
    reduce the total energy by decreasing
    regardless of . Thus, begins to
    increase.
  • Dynamic determination of ?
  • When ? is zero, is determined irrespective
    of the previous submapping and the present
    submapping can be elastically deformed and
    becomes too distorted.
  • When ? is very large, is completely
    determined by the previous submapping i.e., the
    submappings are very stiff, and the pixels are
    mapped to the same locations.

14
Implementation
  • The values of f (m,s)(i,j) (m0,,n) that satisfy
    BC are chosen from the candidates (k,l) by
    awarding a penalty to the candidates that violate
    BC.
  • In the implementation
  • The penalty for violating the third condition ?2
  • The penalty for violating the first and second
    condition ?100000
  • If z-component of W is less than 0, the candidate
    is awarded a penalty by multiplying by ?.

15
Problems
  • Because of bijectivity condition, the pixels on
    the borders of the source image are mapped to the
    pixels on the borders of the destination image.
    This causes artifacts near the borders of the
    resulting images.
  • When there are bifurcations of regions in the
    images, it is useful to abandon BC at bifurcation
    points.
Write a Comment
User Comments (0)
About PowerShow.com