Classifying Kung-Fu Side kicks

zoomzurichΤεχνίτη Νοημοσύνη και Ρομποτική

16 Οκτ 2013 (πριν από 3 χρόνια και 10 μήνες)

95 εμφανίσεις

Classifying Kung
-
Fu Side kicks

With low cost hardware and open source software

Victoria
Værnø

School
of Computer Science & Engineering, Seoul National
University

victoriavarno@gmail.com


Method

Iterations & Results


References


Data

Biointelligence

Lab, Seoul National University, Seoul 151
-
744, Korea
(http://bi.snu.ac.kr)


Background


Research Questions


Motion

capture

technology

and

machine

learning

workbenches

are

accessible

to

the

greater

public,

not

just

computer

scientists

anymore
.


Amateurs

and

professionals

are

responding

to

the

need

for

open

source

technology,

but

links

between

cheap

hardware

(like

web
-
cams

and

Kinect
)

and

open

source

software

is

still

hard

to

find

for

motion

capture
.


This

project

documents

a

basic

experiment

connecting

Kinect

with

open

source

software

and

machine

learning

theory
.

For

a

version

of

the

multilayer

perceptron

algorithm

over

my

motion

capture

data
:


For

separately

created

unseen

data,

what

accuracy

and

“bad”
-
class

label

precision

can

be

obtained?


Are

these

results

judged

good

enough

to

use

in

a

beta

production

of

an

automatic

feedback

application

for

amateur

Kung
-
Fu

training

purposes?


What

challenges

follow

a

relatively

limited

training

set

and

what

basic

machine

learning

techniques

can

reduce

these?

Iterations

of

experiment

and

analysis

to

acquire

deeper

understanding

of

the

data

and

find

the

best

machine

learning

technique

to

make

a

predictor

based

on

the

data

properties
:


Capture

the

kick
-
data

and

identify

the

class

for

each

kick
.


Run

iterations

of




Model

the

data

with

parameter
-
tuned

machine

learning

technique

in

Weka



Test

on

unseen

data

and

analyze



Change

data

and/or

machine

learning

technique

Main

focus

on

the

Multilayered

Perceptron

algorithm

with

variations
.

Meta
-

Learner



Clustering

Result
-

and Data

Analysis

Set 1
:
T
raining data
with

kicks by Victoria.
Set 2
: Test
set

with

kicks by Victoria in di
fferent
circumstances

than

the

training
data.
Set
3
:
Kicks by
the

man Svenn
.

Accurac
y

97%


Precision

o
f

”Bad


1


MLP

-

10
-
fold cross
validation

Yes
,

Unbal
-
anced
.


Blog

with

guides
to
getting

started

on

the

more
advanced

features

of

Weka
:


http
://ianma.wordpress.com/category/weka
/


Witten

I.
H
. &
Eibe

F. & Hall
M. A.

(2011). Data
mining

:
practical

machine

learning

tools

and
techniques
.


3rd ed. Chapter 8,11.
Morgan
Kaufmann

Publishers


Mitchell T.M. (1997). Machine Learning. Chapter 1,3,4,5,8.

McGraw
-
Hill
Science
/Engineering
/Math


Nilsson N. J. (2009). The Quest for Artificial Intelligence.
Cambridge University Press.

Can
be
found

free at
http
://ai.stanford.edu/~nilsson/QAI/
qai.pdf


Weka homepage
:
http://www.cs.waikato.ac.nz/ml/weka
/


Help with code from the open
source community:
http://stackoverflow.com
/

Accuracy

84
%


Precision

o
f

”Bad


1


T
e
s
t

Average
,

mean

a
ccuracy

89
%

Precision

o
f

”Bad


1


After

only

half
of

Data Set 3
was

added

to
the

training
set
,
prediction

accuracy

of

the

model

on

Data
Set 2
increased

by 5
-
9%. The
boosting

algorithm

typically

increased

t
he

results

of

it’s

base
classifier

by 2
-
3%.

Lessons

learned


It is
very

hard for
one

person to
create

unbalanced

motion data.


Using a
boosting

stratagy

to
combat

unbalanced

motion
capture

data
does

have

some

positive
effect
,
but

adding

a different
person’s

motion is far more
efficient
.

”Spend time
gathering

more data
rather

than

tuning a
particular

method
” Nilsson N.J


These

results

are

promising

for
further

investigation

in
machine

learning

for
motion
capturing

with

low

cost

hardware and
open

source

software
.
However
,
the

unseen

test
case
is by a person
also

represented

in
the

training data.
Classifying

unseen

people’s

kicks
remain

unexplored
,
but

light

experimentation

suggests

that

adding

just a
few

kicks
by
new

people

to
the

training data
greatly

increases

the

model’s

generalizability
.

Data
attributes
:

18
joints * 3
dimensions

* 6
frames

per
movie

+ 1
class

lable

= 325
attributes

Clustering K
-
means

K=3