Detection of Curvilinear Regions

1
Detection of Curvilinear Regions
J. Miteran, J. Matas March 2004

• Context image retrieval / object recognition /
  matching

2
Introduction Real world
Presence of numerous objects which can be built
using cylinders or generalized cylinders
Legs of chairs, lamp support, cables
3
Real world
Trees, Pipes
4
Real world
?
5
Real world
Plant, leafs
6
The pins are not detected using extrema
regions However, these regions could be useful
for matching
This is possible using geometrical properties of
the original objects
7
Objective - Theory

• Objective automatic extraction in a viewpoint
  andillumination invariant manner regions from
  real images called curvilinear regions.
• Theory coming from Biederman (Recognition-by-compo
  nents)
• Objects are build from 3-D volumetric primitives
  called Geons (most geons are generalized
  cylinders).
• Objects can be recognized using these primitives
  and their interrelations.

8
Sample of profile
9
Pseudo 3D
10
Sample of profile
Curvilinear profile (crossection)
Curvilinear region
11
Curvilinear regions
Profile
12
Curvilinear regions
Maximum of Gradient
Similar areas
13
Cost function

• The output is a set S of N curvilinear regions.

14
Possible constraints
15
Previous work

• Steger
• An Unbiased Detector of Curvilinear Structures
  in IEEE Transactions on Pattern Analysis and
  Machine Intelligence, 20(2), February 1998,
  113-125
• Unbiased Extraction of Curvilinear Structures
  from 2D and 3D Images Dissertation, Fakultät für
  Informatik, Technische Universität München, 1998
  Herbert Utz Verlag, München, ISBN 3-89675-346-0
• Ziou- Deschenes
• Djemel Ziou Optimal Line Detector. ICPR 2000
  3534-3537
• François Deschènes, Djemel Ziou Detection of
  Line Junctions in Gray-Level Images. ICPR 2000
  3762-3765
• Mouret-Géraud
• Géraud, T. (2003). Segmentation of curvilinear
  objects using a watershedbased curve adjacency
  graph. In Proc. of IbPRIA 2003.
• Mouret, J.-B. (2003). Curvilinear object
  extraction. Technical report, LRDE (EPITA
  Research and Development Laboratory).

16
Steger Objectives
17
Steger

• Main ideas
• filtering using first and second derivatives of
  gaussians (similar to Marr-Hildreth filtering)
• Subpixel localization of zeros
• Linking algorithm
• Determination of line width using edges and
  Taylor polynomial
• Removes the bias of asymmetrical lines
• Uses only constraints 1 and 4 (Boundaries are
  located in local maximum of gradient, the width
  is constant)

18
Steger Results
19
Steger Results
20
Ziou-Deschenes

• Main ideas
• Define an optimal line detector using a line
  model and the 3 Cannys criterions
• Implement the detector using IIR filter ,
  separable case and recursive implementation
  faster than Steger filtering method
• Use classical non maxima suppression as final
  step

21
Ziou-Deschenes Results
22
Ziou -DeschenesResults
23
Mouret-Géraud - Objective
24
Mouret-Géraud

• Main ideas
• Watershed Transform applied to an image of line
  candidates (gradient norm, )
• build a curve adjacency graph
• define a Markov Random Field (MRF) and run a
  Markovian relaxation to get a final segmentation.
• Multiscale approach (Mouret)
• Uses hypothesis of continuity, but only for thin
  lines

25
Mouret-Géraud - Results
26
Preliminary experiments
27
Algorithm

• P(i,j) is a candidate for R if bellowing to edges
  obtained after non maximum supression

28
Results
29
Conclusion