automated extraction and parameterization of motions in large data sets siggraph’ 2004 lucas...

Automated Extraction and Parameterization of

Motions in Large Data Sets

SIGGRAPH’ 2004

Lucas Kovar, Michael Gleicher

University of Wisconsin-Madison

CAIG/CS/NCTU 2

Outline

Introduction

Searching for Motions

Parameterizing Motion

Results & Discussion

CAIG/CS/NCTU 3

Introduction

GoalFinding similar motion segments in a data set and using them to construct parameterized motions

CAIG/CS/NCTU 4

Introduction (Cont.)

HowSearching “Similar” Motion Data Sets

• Multi-step search

• Using time correspondences to determine similarity

• Interactivity through precomputation(match web)

Creating Parameterized Motions• User-specified function F maps blend weights to mo

tion parameters, actually we want F¯¹

CAIG/CS/NCTU 5

Searching for Motions (Cont.)Determine similarity

Corresponding frames should have similar skeleton poses

Frame correspondences should be easy to identify

Time alignment

Monotonically increasing

Continuous

Non-degenerate

CAIG/CS/NCTU 6

Searching for Motions (Cont.)Cell(i, j) : d(M1(ti), M2(tj))

Find the avg and compare against a user-specified threshold €

1D minima

CAIG/CS/NCTU 7

Searching for Motions (Cont.)

D(F1, F2) : distance between two frames of motion( Kovar SCA 2003)

CAIG/CS/NCTU 8

Match Webs Looking for chains of 1D minima

Remove chains below a threshold length

Connecting chains as long as the connecting path is inside the valid region and has a length less than a threshold L

Valid region: extend local minima

CAIG/CS/NCTU 9

CAIG/CS/NCTU 10

Searching With Match WebsMatch sequence

Remove whose avg cell value if greater than € and remove redundant

CAIG/CS/NCTU 11

Searching With Match Webs

Match graph

Node: motion segments

Edge: time alignment

CAIG/CS/NCTU 12

Parameterizing Motion

F: maps a set of blend weights w to a parameter vector p

What we want: a set of parameters => blend weights that produce the corresponding motion

Not guaranteed to be dense or uniform => generate blends to create additional samples

CAIG/CS/NCTU 13

Parameterizing Motion (Cont.)

Motion registration

Sampling strategy

Fast interpolation that preserves constraints

CAIG/CS/NCTU 14

Registration

Timewarp curve s(u)

Ne example motions => each point on s is an Ne-dimensional vector

Automatic determination may fail for more distant motions => identify the shortest path from Mq to every other motion in the match graph

CAIG/CS/NCTU 15

SamplingProduce a dense sampling of parameter space to fill the gaps

Compute the parameters of each example motionCompute a bounding boxRandomly sample points in this region

CAIG/CS/NCTU 16

Interpolation

Given a new set of parameters , to find blend weights

D(): distance between two parameters

CAIG/CS/NCTU 17

Interpolation (Cont.)

Parameters that are not attainable are projected onto the accessible region of parameter space

CAIG/CS/NCTU 18

Results and Discussion

Future worksThe development of alternatives to match webs that are more efficient

Developing methods to ease the data requirements while preserving motion quality

Construct more parameterized motion, ex: leaping motion