Title: Geometric Algorithms in 3D Real-Time Rendering and Facial Expression Modeling (Parts II and III)
1Geometric Algorithms in 3D Real-Time Rendering
and Facial Expression Modeling (Parts II and III)
- Prof. Marina L. Gavrilova
- University of Calgary
2Where we are
- Brief Overview of the Area
- Geometric Approaches for mesh reconstruction and
image morphing - Real-Time Optimally Adapting Mesh Algorithms
- Non Photo-realistic Facial Expression Modeling
and Morphing Algorithms - Swarm Real-time Visualization System
3Part II
4Terminology
- DT Delaunay triangulation
- VD Voronoi Diagram
- FT Feature Transform
- DT Distance Transform
5Preliminaries
- Transformation from a photograph into an artistic
work -
Li and Huang 03
6Preliminaries
- Transformation from an Excel into Artistic Work ?
7NPR vs. PR philosophy
- PHOTOREALISTIC
- GOALS imitate as realistically as possible
facial expression and details - Methodology
- complex geometry fitting,
- generating high density meshes,
- recovering camera positions,
- utilizing data interpolation,
- simulating light conditions,
- applying texturing and other techniques
- Outcome
- Very costly processing, highly realistic
visualization
- NON-PHOTOREALISTIC
- GOALS imitate emotions as closely as possible,
while at the same time allowing the model its own
identity and artistic representation - Methodology
- Utilize artistic drawings
- Feature line extraction
- Expression morphing
- Model partitioning onto regions
- Outcome
- Expression and emotion preservation, efficient
storage, fast processing, artistic visualization
8Animation Trivia
- Face models used in Pixars Toy Story had
approximately how many control points each? - 100
- 1000
- 10000
- None they were made by hand out of clay!
9Animation Trivia
- TOY STORY" used how many SUN workstations to
render images? - 1
- 100
- 1000
- None they were all IBM workstations!
10Animation Trivia
- Pixar have received how many Academy Awards
from the Academy of Motion Picture Arts and
Sciences, including an Oscar for Best Short
Animated Film for 'Tin Toy' and a Technical
Achievement Award for Pixar's RenderMan
software? - 2
- 6
- 12
- One for every animated character created!
11NPR vs. PR
- PHOTOREALISTIC - Final Fantasy (from Final
Fantasy website)
NON-PHOTOREALISTIC - Shrek 10 to 1 preference
Crawford (Shrek movie web site)
12Objectives of our Research
- To develop efficient and realistic portrait
generation method from 2D photographs - To develop fast and easy 2D morphing algorithm
- To create appropriate 3D mesh representation for
model morphing
13Developed Methodology
- Proposed to use NPR modeling by example for
portrait generation - Applied binarization method to produce black and
white images from photograph - Developed segmentation technique for detailed
feature extraction - Used combined edge eextraction strategy
- Used application-dependent style for rendering
- Developed Euclidean distance transform based
algorithm for morphing
14Facial Modeling Research
Lin and Hong 02
Li and Kobatake 97
15Facial Modeling Research
Lin and Hong 02
Gooch 04
Hong 04
16Motivation
- Shortcoming of some previous works
- Facial creases, wrinkles and freckles are hard to
represent - Detail feature lines of different expressions and
which lines? - Efficient morphing
17Contribution
- Task Generate a smooth line-drawing animation
from two input photographs - Input/Output two human face images as input and
output of a NPR facial animation - Methodology
- Create NPR style portrait from a frontal facial
photograph to generate a digital artwork. - Use segmentation and tracking methods to map
expressive lines representing expressive
wrinkles and creases into the image. - Utilize morphing algorithm based on Euclidean
distance transform that does not require control
points or control lines, and finds pixel
correspondence automatically. - Use the developed system as a prototype for 3D
morphing based on controllable mesh deformations
utilizing Delaunay triangulation.
18System FlowChart
19Methodology Use artists rendering as a template
Six Primary Facial Expressions drawn by
artistsanger, disgust, fear, happiness, sadness
and surprise
20Methodology Portrait generator
- Producing black-and-white images using
binarization
21Methodology Details Extraction
Enlarge certain areas
Edge Detection
Delete normal contours
Detail feature lines tracking
Save feature Lines information
22Methodology region subdivision
- Five interested sub-regions are used to extract
details -
23Methodology edge detection
- Edge Detection Process Sobel, Prewitt, Roberts,
LoG, Canny operators (from (b) to (f)), g
combination of e and f
24Methodology details extraction
- Close-up Details Extraction feature
-
25Methodology feature line tracking
26Methodology feature line tracking
27Methodology feature line tracking
- Data Structure to store the feature lines
28Methodology Distance transform for Morphing
Image
Template
29Distance Transform
Given an n x m binary image I of white and black
pixels, the distance transform of I is a map that
assigns to each pixel the distance to the nearest
black pixel (a feature).
30Feature Transform
The feature transform of I is a map that assigns
to each pixel the feature that is nearest to it.
31A Fast Algorithm for Computing the Euclidean
Distance Transform
- Works in Euclidean Metric
- Optimal linear in the number of pixels
- Proceed in two passes over the image, top down
and bottom up - For each row
- Maintain a polygonal chain - a set of pixels in
which the nearest feature may lies - Updates the chain for next row by pruning - it
becomes the minimal set of pixels - Prune Endpoints and Internal Vertices
32Chain Definition
- One Chain for the row
- Contains at most one pixel per column
- Contains the lowest pixel in the column
- After pruning, it contains only those points that
will be nearest features to some points in the
row - Dynamically updated from row to row
33Example of a a Chain
34Pruning Endpoints
- Identifying superfluous endpoints
35Pruning Internal Points
36Pruning Internal Points
37Results on NPR Rendering
38Experimental results for distance transform
algorithm
Algm 1 with pruning was tested Saturation levels
5, 15, 30 Results Linear Running Time,
Regardless of Saturation ( of black pixels in
image)
39Morphing Direction
40Results of utilizing Distance Transform for
Morphing
Starting Frame
Ending Frame
41Results of feature lines extraction
42- Line Simplification and Rendering Results
43- Line Simplification and Rendering
44Animation
453D Facial Model Synthesis
46Brief Preliminaries
- Over the last 30 years, thousands of papers were
devoted to 3D face animation and modeling. - Different classification and registration
techniques have been developed, including those
based on 2D feature matching, multiple point of
view registration (using calibrated cameras),
nearest-neighbor feature comparison,
segmentation-based feature extraction, ICP
(Iterative Closest Point-based Matching), TPS
(Thin Plates Spline) algorithms and others.
473D Facial System Classification(by Akarun,
Gokberk, Salah)
483D Facial System Classification(by Akarun,
Gokberk, Salah) cont.
493D Facial System Classification(by Akarun,
Gokberk, Salah) cont.
503D Facial System Classification(by Akarun,
Gokberk, Salah) cont.
51Voronoi Diagram for 3D face modeling
- Voronoi Diagram is a partitioning of the space
into convex regions such that any point within
the region is closer to the site that originates
this region than to any oethr site in the space. - Delaunay Triangulation is the dual structure that
connects those sites whose Voronoi Regions share
an edge (in 2D) or a face (in 3D).
52Methodology
- Utilize geometric concept Voronoi diagram in
application to FFD (free-form deformation) model.
- Develop a method for 3D facial model synthesis
that combines the traditional free-form
deformation (FFD) model with techniques of data
interpolation based on Voronoi diagrams. - With 18 feature points extracted from 2D facial
images in two orthogonal views, Dirichlet Free
Form Deformation (DFFD) is utilized for modifying
a generic 3D face to produce the individual face.
- The main advantages of this approach over former
extensions of FFD is in removing the constraints
on control lattice and control points location. - By assigning different weights to those control
points, we make the method more adaptable to the
facial deformation. The reconstructed 3D faces
can be used to generate different facial
animations.
53Methodology
- Build DFFD model
- Compute Sibson coordinates using space
partitioning introduced by Voronoi diagram - Compute deformation in 3D using weighted DFFD
relationship based on computed Sibson coordinates - Adapt approach for face synthesis using Delaunay
triangulation
54Illustration
55Illustration
- FaceGen generic mesh and texturing result
56Illustration
- Original face (left) and generated face (right)
57Extensions
- The developed methodology is currently being
utilized for 3D face morphing and expression
synthesis. -
58Part III Strategic Modeling
59Problem Definition
- Given a set of unsophisticated agents and their
behavior. - Examples ants, bees, fish, animals, robots
- Complex behavior can emerge as a result of
collective intelligence
60Our Methodology
- Stigmergy (approach where agents interact with
environment) is usually realized through
neighborhood cohesion (Boids flocking) or through
chemical traces modeling. - We introduce a novel behavioral Genotype encoding
with novel agent Classification Genetic Algorithm
and more complex set of rules determining missile
as well as target behavior. - We also test it in a complex 3D environment.
61Illustration Missile Basic Maneuvers
XF
F
U
N
LU
A
62Rendering and Physical Engine
- Regular physics engine will not suffice
- Approximation aggravates trajectory computation
- Construct original visualization system
- Advanced look-ahead estimation based physics
engine - Robust Rendering engine
- Anisotropic Texture filtering
- Multiple LOD based geometry rendering
- Particle engine
- Highly optimized exclusive adaptive pattern
identification - Flexible and robust system
63The Simulation Engine
- Robust design Separation of Rendering modules
from the simulation - Implement Command Console
- Runtime performance is highly efficient
- For 50 missiles
- Full quality rendering at 50FPS (better than most
oethr results reported) - Simulation runs up to 50 times faster (FPS2200)
if rendering is turned off (for evolutionary
algorithm) - Excellent Rendering quality (anisotropic texture
mapping, particle engine are utilized)
64More Results on Strategic Modeling
65Conclusions
- Presentation discussed two systems developed for
real-time massive 3D terrain rendering and for
NPR facial expression modeling. In both cases,
efficient and robust results are achieved through
the utilization of computational geometry
algorithms. - It is utilized in other practical 3D modeling
applications - collision detection optimization project in
collaboration with Prof. Ivana Kolingerova,
Computational Geometry Group (University of West
Bohemia, Czech Republic), - protein modeling project carried out with Prof.
Sergei Bereg (University of Texas at Dallas,
USA), - material structure modeling project together with
Prof. Nikolai Medvedev, Novosibirsk State
University, Russia.
66 Acknowledgements NSERCGranting Agency,
CanadaGEOIDENetwork of Excellence,
CanadaCanadian Foundation for Innovation All
students of SPARC LAB and BT Lab, University of
Calgary, in particularly Russel Apu and Luo Yuan
CollaboratorsAudience