Feature-Based Textures

1
Feature-Based Textures

• Ganesh Ramanarayanan
• Kavita Bala
• Bruce Walter
• Cornell University
• EGSR 2004

2
Motivation

• Textures have fixed resolution
• Goal resolution independent texturing
• Improving texture quality when zooming in
• Applications games, interactive walkthroughs

3
Solution

• Feature-Based Textures (FBTs)
• Humans are sensitive to features boundaries in
  the texture with sharp contrast
• Store features as resolution-independent lines /
  curves
• Interpolate from reachable samples those on the
  same side of all features

4
Quality
Using standard texture map
5
Quality
Using Feature-Based Texture
6
Related Work

• Vector formats SVG, Postscript
• Procedural textures Ebert94
• Superresolution Huang84, Elad97, Borman98
• Discontinuity detection and use
• Vision Canny87, Perona90, Malik01
• Point data sets Pauly03
• Visibility events Drettakis94, Durand99,
  Duguet02
• Discontinuity meshing Heckbert92,
  Lischinski92
• Silhouette clipping Sander00

7
Related Work

• Discontinuity-based 2D representations
• NPR Salisbury96
• Edge-and-point interactive rendering Bala03
• Silhouette shadow maps Sen03
• Concurrent research
• Bixels Tumblin04
• Silmap textures Sen04

8
Outline of Talk

• Overview
• FBT Usage
• FBT Representation
• Results and Discussion

9
Overview

• Each FBT texel stores samples and features
• One sample in each texel region
• Most of the FBT works like a standard texture map
• Only do more computation near features

feature
samples

10
Algorithm
point
Image plane
11
Algorithm

1. Map 3D point to 2D texture point p

12
Algorithm

1. Map 3D point to 2D texture point p
2. Find the texel containing p

13
Algorithm

1. Map 3D point to 2D texture point p
2. Find the texel containing p
3. Find the texel region containing p

14
Algorithm

1. Map 3D point to 2D texture point p
2. Find the texel containing p
3. Find the texel region R containing p
4. Look up reachable samples

15
Algorithm

1. Map 3D point to 2D texture point p
2. Find the texel containing p
3. Find the texel region R containing p
4. Look up reachable samples
5. Bilinearly interpolate and return result

16
Outline of Talk

• Overview
• FBT Usage
• FBT Representation
• Results and Discussion

17
Step 3 Region Finding

• How do we find which of these four regions a
  point lies in?

18
Step 3 Region Finding

• Simple feature divides texel into two regions
• Distinguish these regions with a single ray
  intersection parity test against feature
• Cast towards side with no feature intersection

0
1
19
Step 3 Region Testing

• n simple features divide area into n1 regions
• Identify region using linear search

20
Step 3 Region Testing

• n simple features divide area into n1 regions
• Identify region using linear search

21
Step 3 Region Testing

• n simple features divide area into n1 regions
• Identify region using linear search

22
Step 3 Region Testing

• n simple features divide area into n1 regions
• Identify region using linear search

23
Feature Intersections

• When features intersect, parity test for region
  determination can be ambiguous
• Solve by introducing horizontal bands

24
Step 4 Sample Lookup

• If a texel has no features, use ordinary bilinear
  interpolation
• What do we do when there are features?

1
3
1
3
s
2
s
2
25
Step 4 Sample Lookup

• Sample in lower left region of texel
  representative sample
• Interpolate using regions sample s and reachable
  representatives

1
3
1
1
s
2
s
s
26
Step 4 Sample Lookup

• Sample in lower left region of texel
  representative sample
• Interpolate using regions sample s and reachable
  representatives

1
3
3
3
s
2
s
3
27
Outline of Talk

• Overview
• FBT Usage
• FBT Representation
• Results and Discussion

28
Creating FBTs

• Input samples, features, and FBT resolution
• Finding features
• Automatic extraction (vectors, feature detection,
  tracing)
• Manual specification (from scratch or extracted
  features)

Tracing
Vectors
User-drawn
29
Resolution Tradeoff

• More FBT texels means
• More memory usage
• More texels with no features for faster lookup
• Denser sampling of input for better gradients

30
Sample Invalidation

• Samples close to a feature are prefiltered
• Such samples should be eliminated

31
Reachability Graph

• Decompose space into sub-regions

32
Reachability Graph

• Decompose space into sub-regions
• Divide texel at all feature/feature
  intersections, feature/texel intersections, curve
  maxima/minima

33
Reachability Graph

• Decompose space into sub-regions
• Divide texel at all feature/feature
  intersections, feature/texel intersections,
  spline maxima/minima
• Use sub-regions to form reachability graph

34
Hole Filling

• Fill holes with closest available samples

35
Hole Filling

• Fill holes with closest available samples
• Collapse sub-regions into main regions

36
Outline of Talk

• Overview
• FBT Usage
• FBT Representation
• Results and Discussion

37
Example Inputs
Stop sign (vector)
Yin yang (vector)
38
Example Inputs
Stained glass (user)
Flower (user)
Wizard skin (user)
39
Results
SVG Vector Format
FBT 230x256 416 KB
40
Results
FBT 230x256 416 KB
FBT 16x16 9 KB
Bilinear 460x512 690 KB
Bilinear 64x64 12 KB
41
Results
FBT 128x128 96 KB
FBT 256x256 434 KB
Bilinear 128x128 48 KB
Bilinear 256x256 192 KB
42
Results 3D Models
FBT 230x256 416 KB
Bilinear 460x512 690 KB
43
Results 3D Models
FBT 230x256 416 KB
Bilinear 460x512 690 KB
44
Results 3D Models
FBT 230x256 416 KB
Bilinear 460x512 690 KB
45
Results 3D Models
Artwork from Warcraft III Reign of Chaos
provided courtesy of Blizzard Entertainment
FBT 256x256 357 KB
Bilinear 256x256 192 KB
46
Results 3D Models
Artwork from Warcraft III Reign of Chaos
provided courtesy of Blizzard Entertainment
FBT 256x256 357 KB
Bilinear 256x256 192 KB
47
Results 3D Models
Artwork from Warcraft III Reign of Chaos
provided courtesy of Blizzard Entertainment
FBT 256x256 357 KB
Bilinear 256x256 192 KB
48
FBT Properties

• 90 of texels No features
• Most texels similar to standard texture maps
• Low amortized lookup cost
• Low amortized storage cost
• 99 of texels lt 2 features
• Fixed-size representation possible for GPUs
• Promising results on NV40 even without using new
  functionalities (early-out, etc)

49
Conclusions

• FBTs combine samples with resolution-independent
  features
• Flexible representation encompassing vector and
  raster image formats
• Applications games, interactive walkthroughs
• Future work
• MIP-mapping and antialiasing
• Handling a wider variety of textures

50
Questions

• graman_at_cs.cornell.edu
• kb_at_cs.cornell.edu
• bjw_at_graphics.cornell.edu
• FBT Webpage
• http//www.cs.cornell.edu/graman/papers/egsr04fbt