Specifying Label Layout Styles by Example

1
Specifying Label Layout Styles by Example

• Ian Vollick1, Daniel Vogel1,
• Maneesh Agrawala2, Aaron Hertzmann1
• 1University of Toronto, 2UC Berkeley

2
Romanes, G.J., "Cunningham's Manual of Practical
Anatomy 14th Edition." Oxford University Press
(1977)
3
What is Label Layout?

• The label layout problem consists of
• An arrangement of the labels is called a layout.

manoa
nuuanu
kipapa
4
Bad Layouts

• Good layouts do not have
• overlapping elements
• elements outside page boundary

5
Stylistic Decisions
6
Motivation
Left Sobotta, Johannes, "Atlas and text-book of
human anatomy, Philadelphia, Saunders (1914)
Right Pratt, Neal E., "Clinical Musculoskeletal
Anatomy," J.B. Lippincott Company (1991)
7
Motivation
Swindler, Daris Ray, "An atlas of primate gross
anatomy." University of Washington Press (1974)
8
Motivation

• Many stylized layouts are needed for
• Automotive manuals
• Anatomical atlases
• Catalogues, etc..

http//www.longenterprises.com
9
Automated Layout Systems

• Rule-based layout
• e.g. Visio, Interface Builder
• Energy-based layout
• e.g. LineDrive, GADGET

10
Energy-based Layout

• Goal find XS arg minX E(X)
• E(X) ?i Ei(X)
• Ei(X) are objective terms

11
Objective Terms

• Eoverlap

nuuanu
kipapa
manoa
12
Objective Terms
Egap
Edist
Epref?
Ealign
Eexitf
13
Objective Terms Edist

• Edist (X ?) Sj dj

14
Objective Terms Eoverlap

• Eoverlap (X ?) Siltj f(-g di,j)

15
Parameterized energy function

• E(X ?) Si ai Ei(X ?i)
• ? a1 , ? 1 , a2 , ? 2 ...

16
Problem

• How do we set ??

17
Previous Work

• Preference Elicitation
• e.g. Masui 1994, Gajos Weld 2005

18
NIO

• Nonlinear Inverse Optimization
• (Liu, Hertzmann Popovic, 2005)

19
Problem statement

• Given an optimal, artist-created layout, what was
  the energy function?
• Given
• layout XT
• layout problem
• Determine ?

E(X,?)
XT
20
Obvious Approach

• The obvious approach doesnt work

XT arg min E(X q) 2
21
Idea

• Goal
• E(XT ?) minX E(X ?)
• Learning objective function
• G(?) E(XT ?) - minX E(X ?)

22
Learning

• XS(?) arg minX E(X ?)
• G(?) E(XT ?) - E(XS(?) ?)

-
-
use gradient descent with line search
23
Learning
E(X ?)
XS
XT
XS
24
Validation
ground truth
estimated
Synthetic
25
Validation
26
Results
Horizontal/Vertical
27
Results
Horizontal Groups
28
Results
Left and Right groups, horizontal leaders
29
Results
Left and Right Groups
30
Results
Parallel Leader Lines
31
Limitations

• The example must be sufficiently complex to
  constrain q.
• The style must lie in the space of styles S
  q1,q2,q3,...

manoa
32
Limitations
Horizontal
33
Summary and Future Work

• Learning layout style from a single example
• General framework for learning style
• Future work
• Improve performance
• Interactive workflow

34
Thank-you
35
Alpha Constraints

• Constraints ?j ?j 1, ?j gt 0
• to prevent E0 everywhere