a dynamic noise primitive for coherent stylization, egsr 2010
DESCRIPTION
Presentation of our paper called "A Dynamic Noise Primitive for Coherent Stylization" published at EGSR 2010TRANSCRIPT
A DYNAMIC NOISE PRIMITIVE
FOR COHERENT STYLIZATION
Pierre BénardJoëlle ThollotGrenoble Universities / INRIA Rhône-Alpes
Ares LagaeKatholieke Universiteit LeuvenREVES - INRIA Sophia-Antipolis
Peter VangorpGeorge DrettakisREVES - INRIA Sophia-Antipolis
Sylvain LefebvreALICE - INRIA Nancy / Loria
Stylization of 3D Animations• 3D scene 2D appearance
2
Stylization of 3D Animations• 3D scene 2D appearance
• Stylized color regions– 2D medium: a pattern– Temporal coherence
3
Paper
Pencil strokes
Watercolor pigments
Paint strokes
Hand –made animation« Il pleut bergère », Jérém
y Depuydt (2005)
4
Popping Temporal continuity
Naïve CG solutions
5
Shower-door effect Coherent Motion
Traditional mapping Flatness
Temporal Coherence Problem• Extreme cases Requirements
6
Coherent motion Temporal continuity
Flatness
Traditional mapping
Shower-doorPopping
Contradictory requirements:solution find a compromise
3 goals to ensure at best
• Additional challenges– Flexibility variety of styles– Interactivity artistic control– Evaluation quality of the trade-off
Coherent motion Temporal continuity
Flatness
7
PREVIOUS WORK
Texture -Based methods• Object-space Blending artifacts Perspective distortion
Coherent motion
Temporal continuity
Flatness
Traditional mapping
Shower-doorPopping
Object-space texture mapping[KLK*00,PHWF01,FMS01,BBT09]
[BBT09]
9
Texture -Based methods• Object-space• Screen-space Sliding
Coherent motion
Temporal continuity
Flatness
Popping
[CTP*03]
Screen-space texture mapping[CTP*03,CDH06,BSM*07,BNTS07]
Object-space texture mapping[KLK*00,PHWF01,FMS01,BBT09] 10
Shower-door
Few -Primitive methods Clutter / holes Popping
11
Coherent motion
Temporal continuity
Flatness
Popping
[Mei96]
Few-primitive methods [Mei96,Dan99,HE04,VBTS07]
Object-space texture mapping[KLK*00,PHWF01,FMS01,BBT09]
Screen-space texture mapping[CTP*03,CDH06,BSM*07,BNTS07]
or
Few -Primitive methods
12
Vanderhaeghe et al. EGSR 2007
Key Insight• Blending a large number of primitives
– Reduce popping artifacts– Individual primitives merge texture
13
Many-Primitive methods
Coherent motion
Temporal continuity
Flatness
[KC05]
Many-primitive methods[KC05,BKTS06]
14
Object-space texture mapping[KLK*00,PHWF01,FMS01,BBT09]
Few-primitive methods [Mei96,Dan99,HE04,VBTS07]
Screen-space texture mapping[CTP*03,CDH06,BSM*07,BNTS07]
NPR GABOR NOISE
Procedural noises• Sparse convolution [Lewis 84,89]• Spot Noise [van Wijk 91]• Gabor Noise [LLDD09]
Our trade-off: NPR Gabor Noise
16
Gabor Noise [LLDD09]• Offers significant spectral control• Support anisotropy• Is fast to evaluate
17
See “State of the Art in Procedural Noise Functions”, EG 2010 for comparisons with previous work
Gabor Noise [LLDD09]• Definition
Sum of randomly positioned and weighted kernels
18
Gabor kernel noiserandom positionsand weights
NPR Gabor Noise• Basic principles follow from the goals
– Flatness Noise parameters in 2D screen space Evaluation in 2D screen space
19
2D Gabor noise [LLDD09]
NPR Gabor Noise• Basic principles follow from the goals
– Flatness– Coherent motion
Point distribution on the surface of the 3D model
20
Surface Gabor noise [LLDD09]
2D Gabor noise [LLDD09]
NPR Gabor Noise• Basic principles follow from the goals
– Flatness– Coherent motion
21
Surface Gabor noise [LLDD09]
NPR Gabor noise2D Gabor noise [LLDD09]
NPR Gabor Noise• Basic principles follow from the goals
– Flatness– Coherent motion– Continuity
Smooth LOD mechanism
22
Surface Gabor noise [LLDD09]
NPR Gabor noise2D Gabor noise [LLDD09]
GPU Splatting Algorithm• Sample 3D triangles
– 2D Poisson distribution with constant screen space density
– PRNG: seed = triangle ID
23
Farsmall screen
arealess points
Closelarge screen
areamore points
GPU Splatting Algorithm• Generate 2D point sprites
24
Point distribution Texture sprites
LOD Mechanism• Blending scheme using statistical properties
– Reduce popping– Preserve noise appearance
25
STYLES
Style Design• Standard techniques from procedural texturing
and modeling [EMPPW02]– Threshold
• Smooth step function • X-toon textures [BTM06]
– Compositing (alpha-blending, overlay)– Local control
• Curvature noise orientation• Shading noise frequency
• Interactive feedback
Threshold texture
27
Style Design
28
Results :
29
isotropic as well asanisotropic patterns
30
Results :local variationaccording to shading
Results :
31
local orientation guided by surface curvature
USER STUDY
User Study: Motivation• Evaluate success of various solutions
according to
• Relative importance of these criteria
33
Coherent motion Temporal continuity
Flatness
User Study: Setup• Methodology
– 15 naïve subjects, ~ 20-30 minutes• Ranking tasks
“Rank the images/videos according to … ”
34
User Study: Compared methods
35
Adv D2D DST
ours SD TM
Extreme cases
Local screen-spaceGlobal screen-space Object-space
User Study: Flatness• Simple stimuli
Adv D2D DST
ours SD TM
36
Object-space
User Study: Flatness• Complex stimuli
Adv D2D DST
ours SD TM
37
User Study: Flatness“Rank the images according to
how flat they appear.”• Simple stimuli
– Image-space methods more flat– Object-space methods less flat
• Complex stimuli– High variance confusing question– Many 3D cues flatness not perceived
38
• Simple stimuli
User Study: Dynamic stimuli
39
• Complex stimuli
User Study: Dynamic stimuli
40
User Study: Coherent motion“Rank the videos according to how coherently
the pattern moves with the object.”• Simple stimuli
– Object-space methods more coherent– Shower-door least coherent– Image-space methods provide a tradeoff
• Complex stimuli– Same conclusions– Our approach slightly better than other
image-space methods41
“Rank the videos according to how much the pattern changes over time.”
• Simple stimuli– High variance– Advection and ours produce more changes
“swimming” artifacts
• Complex stimuli– Shower door and D2D least changes– Others perceived equally
User Study: Temporal continuity
42
User Study: Pleasantness“Rank the videos according to how
pleasant you find them in the context of cartoon animation.”
• Complex stimuli– Object-space approaches more pleasant– Our methods first image-space approaches
43
User Study: Pleasantness• Strong correlation with “motion coherence”
most important criteria to preserve
• NPR Gabor noise performs well
44
CONCLUSIONSAND FUTURE WORK
45
User Study• First step toward formal evaluation
– Flatness hard to see in complex scenes– Motion coherence predominant criteria
• Intrinsic limitations– Hatching other styles– Naïve users professional artists
• Objective metric– Statistical texture measures [BTS09]– Optical flow analysis
46
NPR Gabor Noise
• New primitive for coherent stylization
• Interactive scheme: remaining popping “Procedural” evaluation [LLDD09]
– Slower, but should avoid popping– Useful for high quality offline rendering
• Temporally coherent spot noise Additional patterns
47
Thanks !
Styles cookbook and experiment stimuli:
http://artis.inrialpes.fr/Publications/2010/BLVLDT10
Acknowledgments• Laurence Boissieux, Kartic Subr, Adrien Bousseau and Marcio Cabral• The participants of the study• Research Foundation-Flanders (FWO), CREA (K.U.Leuven)
User Study: Flatness• “Rank the images according to how flat
they appear.”– Simple stimuli
– Complex stimuli
50
less flat more flat
DST TM Adv ours SD D2D
less flat more flat
TM SD DST ours D2D Adv
50
User Study: Coherent motion• “Rank the videos according to how coherently
the pattern moves with the object.”– Simple stimuli
– Complex stimuliless coherent more coherent
SD D2D ours Adv DST TM
SD D2D Adv ours TM DST
SD Adv D2D ours TM DST
SD Adv ours D2D TM DST
translate:
rotate:
zoom:
less coherent more coherent
51
Object-space
• “Rank the videos according to how much the pattern changes over time.”– Simple stimuli
– Complex stimuli
User Study: Temporal continuity
52
Adv ours D2D DST TM SD
TM DST ours Adv D2D SD
Adv SD TM DST ours D2D
Adv ours D2D DST SD TM
more change less change
52
translate:
rotate:
zoom:
more change less change
User Study: Pleasantness• “Rank the videos according to how
pleasant you find them in the context of cartoon animation.”– Complex stimuli
80
SD D2D Adv ours DST TM
less pleasant more pleasant
53
Object-space