Geometric clustering for line drawing simplification

1
Geometric clustering for line drawing
simplification

• Pascal Barla Joëlle Thollot François Sillion
• ARTIS, GRAVIR/IMAG-INRIA

2
Introduction

• Line drawings are useful
• Convey shape, tone, style
• Used in illustration, art
• Created in many different ways
• Complexity issues
• Artists know how to tune
• complexity
• Computers dont
• Often too many lines

3
Problem statement

• Lines from various sources
• Scanned drawing
• Digital drawing
• Image processing
• Non-photorealistic rendering
• Simplification
• Smaller set of lines
• Keep drawings overall structure

Scanned drawing of a hand
Non-Photorealistic Rendering Grabli
4
Outline

• Related Work
• Methodology
• Clustering algorithm
• Geometric strategies
• Results
• Conclusions

5
Related work

• Density reduction
• trees in object-space Deussen
• in image-space WilsonGrabli
• Indication
• for complex textures
• Winkenbach
• Oversketching
• smoothed Baudel
• constrained Igarashi

Density reduction Grabli
Texture indication Winkenbach
Oversketching tool Baudel
6
Related work

• levels of detail for NPR
• In texture-space
• Tonal Art Maps Praun
• In object-space
• WYSIWYG NPR Kalnins
• Overall limitations
• Specific solutions
• Simplify remove
• No perceptual consideration

NPR with hatching patterns Exhibiting LOD
behaviors Kalnins
7
Related work

• Perceptual organization BoyerSarkar
• Only group lines
• Based on human perception
• Study criteria independently (e.g., parallelism)

A schematic sun figure and the two largest
parallel groupings Rosin
8
Contributions / limitations

• Contributions
• Identify common behavior
• Oversketching, Density reduction and Levels of
  detail
• Perceptually motivated
• Various simplification strategies
• Not only deletion
• Limitations
• Low-level
• Static 2d drawings

9
Outline

• Related Work
• Methodology
• Clustering algorithm
• Geometric strategies
• Results
• Conclusions

10
Methodology

• Single control param e
• simplification scale
• 2 steps
• Automatic Clustering
• Common to envisioned
• applications
• Line creation
• Geometric strategies
• Application dependent

Clustering
Line creation
11
Methodology

• Input
• 2d Vectorized lines
• Attributes e.g., color
• Static drawings
• Clustering output
• Line clusters
• Final output
• Vectorized lines
• attributes

Clustering
Line creation
12
Methodology

• Proximity is not enough
• Forks
• Hatching groups

Unnatural fork behavior
Two simplifying lines keeping underlying fork
structure
Unnatural hatching group behavior
Three simplifying lines keeping underlying stack
structure and orientation
13
Methodology

• Perceptual grouping Palmer
• Criteria proximity, parallelism, continuation,
• and color.
• Integrated in clustering constraints
• Definition of an e-group (see paper)

14
Outline

• Related Work
• Methodology
• Clustering algorithm
• Geometric Strategies
• Results
• Conclusion

15
Clustering algorithm

• Clustering partition
• Greedy algorithm

16
Clustering algorithm

• Clustering partition
• Greedy algorithm
• Clustering 2 lines/groups
• Do they form an e-group ?
• Error measure
• Using attributes

e
17
Clustering algorithm
e

• Clustering a pair of lines
• Example of an invalid pair
• (pb. with parallelism)

18
Clustering algorithm
e

• Clustering a pair of lines
• Example of an invalid pair
• (pb. with parallelism)
• Five valid configurations (see paper)
• Correspond to e-groups on a pair of lines
• Favor parallelism, continuation and proximity

19
Clustering algorithm

• Error measure
• Based on proximity
• Normalized between 0 and 1

20
Clustering algorithm

• Error measure
• Based on proximity
• Normalized between 0 and 1
• Can take attributes into
• account (e.g. color)
• Normalized between 0 and 1
• Combined in a multiplicative
• way

21
Clustering algorithm

• Implementation
• Clustering graph
• Graph node line
• Graph edge valid pair
• Error stored on edges

e
22
Clustering algorithm

• Implementation
• Clustering graph
• Graph node line
• Graph edge valid pair
• Error stored on edges
• Greedy algo edge collapse
• Collapse min error edge
• Delete degenerated edges
• Update graph locally

e
23
Outline

• Related Work
• Methodology
• Clustering algorithm
• Geometric Strategies
• Results
• Conclusion

24
Geometric strategies

• Geometric strategies
• Work on clustering output
• May use clustering history
• Many possibilities
• Application dependent

Clusters
25
Geometric strategies

• Geometric strategies
• Work on clustering output
• May use clustering history
• Many possibilities
• Application dependent
• 2 basic strategies
• Average line
• Most significant line

Clusters
Average line strategy
Longest line strategy
26
Outline

• Related Work
• Methodology
• Clustering algorithm
• Geometric Strategies
• Results
• Conclusion

27
Results

• Density reduction (scanned drawing)
• A single strategy
• Average line

357 input lines
87 output clusters
28
Results

• Density reduction
• (3D model)
• Two different strategies
• Average line for the
• leaves
• Longest line for the
• inner part

531 input lines
256 clusters
294 clusters
29
Results

• Density reduction
• (scanned drawing)
• Taking attributes
• into account
• Lab color threshold

30
Results

• Oversketching
• Apply simplification iteratively
• Drawing sensitivity e
• Each new a sketch has its own sensitivity
• Specific average line strategy
• Give higher priority to last drawn line
• See video

31
Results

• Levels of detail

32
Results

• Levels of detail
• Increasing e
• Simplify output of finer level
• Two different strategies
• Average line for contour
• Longest line for hatching
• See video

33
Conclusions
34
Conclusions

• 2-step approach is valuable
• Analysis of common behavior
• Adaptation to application goals
• 3 application examples

35
Conclusions

• 2-step approach is valuable
• Analysis of common behavior
• Adaptation to application goals
• 3 application examples
• Perceptual grouping
• Incorporate a human vision model in NPR
• Perception of a drawing

36
Conclusions

• Future work

37
Conclusions

• Future work
• Improve clustering
• More perceptual criteria (e.g closeness)
• Individual control for each criterion
• Medium- and high-level processing (i.e drawing
  structure)

38
Conclusions

• Future works
• Create new applications
• Automatic creation of Tonal Art Maps
• Morph transitions for LODs
• Clustering of 2d lines for animation
• Simplification of lines lying on surfaces

39
Acknowledgements

• Gilles Debunne for the video
• ARTIS teams many reviewers
• Lee Markosian and Chuck Hansen for english
  cleanup.