SLIDE 1
Artistic Stylization of Face Photos Based on a Single Exemplar - - PowerPoint PPT Presentation
Artistic Stylization of Face Photos Based on a Single Exemplar - - PowerPoint PPT Presentation
Zili Yi Supervisor: Minglun Gong 2015-11-01 Artistic Stylization of Face Photos Based on a Single Exemplar Outline Introduction Background Method Results Conclusion Reference input Introduction example Style Transfer
SLIDE 2
SLIDE 3
Introduction
input example
- utput
Style Transfer Stylization of Faces Non-photorealistic Rendering
Challenge
Region-specific Viewers’ Sensitivity to face Variation of faces
SLIDE 4
Style-specific methods
Pencil drawing, Lu c. et. al. Tessellation, Li h. et. Al. Halftoning, Li H. et. al. Screening, Li H. et. Al.
SLIDE 5
Data driven methods
Guided Transfer, Li H. Style transfer, Shih Y. et. al. Data Driven Cartoonizing, Zhang
- Y. et. al.
SLIDE 6
Our Task Data Driven Style-Specific General stylization with a single Example
SLIDE 7
Background
Face Detection & Alignment Non-Photo Realistic Rendering Color Style Transfer Optimized Texture Synthesis
SLIDE 8
Framework of Our Method
Example Color Transfer Texture Transfer
SLIDE 9
Framework of Our Method
Example Color Transfer Texture Transfer
SLIDE 10
Framework of Our Method
Example Color Transfer Texture Transfer
SLIDE 11
Color Transfer
Example Color Transfer Texture Transfer
SLIDE 12
Color Transfer
Face landmark
label assignment dense correspondence color assignment
face detection& normalization 320*448 image
SLIDE 13
Texture Transfer
Example Color Transfer Texture Transfer
SLIDE 14
Texture Transfer
We view the problem of Texture Transfer as Edge-
Preserved Texture synthesis problem
Optimized Texture synthesis
Iteratively optimizing and voting
SLIDE 15
Optimizing
Finding the Nearest Neighbor Field Generalized PatchMatch
‘ measures the L2 distance Between patch and S E
SLIDE 16
Voting
Simply averaging the Nearest Neighbors
SLIDE 17
Pyramid
The texture is synthesized in a pyramid manner
Coarse to fine 10 scales (ratio 1.5)
SLIDE 18
Preserve the edges
Pyramid Edge Map
Using edge maps as the preserving mask
SLIDE 19
The final synthesized image
SLIDE 20
Experiment Result
SLIDE 21
Results
SLIDE 22
Comparison
SLIDE 23
Comparison
SLIDE 24
Conclusion
Strength
Robust to gender, skin color, hair style, face accessories, beards,
glasses, slight variation in poses, lightening condition
Able to transfer wide varieties of styles (pencil drawing, sand
drawing, oil painting, mosaic, screening, water color painting, Chinese painting and Pyrography)
Limitation
Time consumption (over 40 mins to process one image) Unable to handle stylization which involves geometry
exaggeration
T
end to blur textures with long curve patterns.
SLIDE 25
Reference
SLIDE 26