Latent Doodle Space

1
Latent Doodle Space

• William Baxter1,
• Ken-ichi Anjyo2
• OLM Digital, Inc.
• EUROGRAPHICS,2006

OLM Digital, Inc. EUROGRAPHICS,2006
Presented by C.M. Hsu
2
Abstract
3
Abstract

• Major tech.
• A heuristic algorithm to match strokes between
  the inputs.
• Extract a low dimensional latent doodle space
  from the inputs.

4
Applications
The Randomized Stamp Tool
Input
5
Applications
Handwriting synthesis
Input
Output
6
Outline

• Overview
• Related Work
• Stroke Matching Algorithm
• Building the Latent Doodle Space
• Conclusions and Future Work
• Demo

7
Overview
8
Overview

• 1. Similar to Computer-assisted in-between
  algorithms
• 2. N-way correspondences, not pair
• 3. Competitive-learning algorithm
• K-means-like
• Match stokes based on Kuhn-Munkres method
• O(N3)

3. Reverse the parameterization of strokes to
improve the point-to-point correspondence
4. Build the latent control space
9
Related Work
10
Related Work Stroke Correspondences

• The order of strokes between two images are
  identical, Burtnyk N., SIG75
• Closed shape only, Sederberg, SIG92

11
Related Work A low-dimensional latent space

• Eigenfaces by PCA, Turk, Nerosci91
• Multidimensional motion interpolation by Radial
  basis functions (RBF) Rose et al.,IEEE98
• The Gaussian Process Latent Variable Model
  (GPLVM), Lawrence, NIPS04
• Create keyframe from ex. by GPLVM, Grochow K.,
  CV04

12
Related Work

• Create many drawings from a few ex., Kovar,
  UIST01

13
Stroke Matching Algorithm

• Finding Stroke Correspondences
• The Assignment Cost Matrix
• Stroke Re-sampling and Alignment

14
Finding Stroke Correspondences

• K-means like
• Initialize stroke-to-cluster assignment
• Clustering by the drawn order of strokes simply
• Update the cost matrix
• Reassign stroke based on new clusters
• Linear assignment problem (strokes ? clusters)
• Constrain one stroke per drawing to each cluster
• Kuhn-Munkres algorithm, (N3), N as number of
  strokes
• If reassignments made, goto 2

15
The Assignment Cost Matrix

• Eedecet
• Translation error, ed
• The mean of the stroke differs from the mean of
  the cluster
• Orientation and eccentricity error,ec
• The covariance axe of the stroke differs from
  those of the cluster
• Topological matching cost, et
• The connectivity of the stroke differs from the
  connectivity of the members of the cluster

16
Translation error

• ed The mean of a stroke differs form the mean of
  the cluster

17
Orientation and eccentricity error

• ec The covariance axe of the stroke differing
  from those of the cluster

18
Topological matching cost

• et Connectivity cost of the stroke differing
  from
• the connectivity of the members of the cluster
• Csj the number of strokes ? the ith stroke in
  the same drawing.
• Ccj the number of clusters ? the ith cluster by
  average distance.

19
Stroke Re-sampling and Alignment

• Same number of points on the corresponding
  strokes
• for RBF, Gaussian process regression
• Reverse backward strokes
• The total distance error between two strokes is
  lower when the point ordering is reversed

20
Building the Latent Doodle Space
Feature Vectorm p1,p2,..pn of Strokem
PCA 2 principal components
PCA 2 principal components

• RBF
• Gaussian
• Thin plate spline, r4logr

thin plate spline RBF
GP
best
21
Conclusions and Future Work

• Using machine learning
• Find good assignment weights in cost function
• Ex S.T. like a support-vector classifier could
  be trainned to assign strokes to clusters.
• Allow user to appraise the products form as
  latent space. Kovar, UIST01
• Accept scanned drawing
• Accept completely free-form hand-drawn sketch
  without the line constrain of uniform width.

22
End