tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
the reshuffled sequence as ˆSgˆSRF3P Differ
ent from existing methods we propose a SpatialWeighted
Nuclear Norm SWNN optimization model that imposes
lowrank constraint on the reshuffle of the proxy shapes
SgˆSΛinstead of ˆS so the new lowrank optimization
model can be written as
min
ˆSS 6
The key to the model 6 is how to design the spatial weight
matrix Λ Next we present a method of constructing proxy
shapes Si i 1 Fto relax the lowrank constraint on
severe deformation regions We first introduce an algorithm
for segmenting objects according to the level of deformation
in different regions Based on this a kernel based method
for the construction of proxy shapes is proposed
332 Nearly Rigid Region Segmentation
In reality not all regions of nonrigid objects undergo the
same deformations but are spatiallyvariant For example
a talking mouth on a human face will undergo more severe
deformation than the nose Thus different regions should
fit the lowrank regularization differently In this section
we introduce the partition algorithm for different regions of
nonrigid objects see Fig 2a for the process
We define Tjˆ stjR31F
t1as the jth column of
shape sequence ˆS which represents a 3D trajectory The
xyz coordinates of a 3D trajectory can be viewed as time
domain signals and can be transformed into the frequency
domain by Discrete Fourier Transform DFT
dkF F12PF
t1ˆ stje2πitkF k 0 F17
we define ωkkF which symbolize the different fre
quency components Intuitively when the time domain sig
nal changes drastically its corresponding frequency domain
signal will contain more highfrequency components We
can compare the severity of different time domain signal
changes by calculating the intensity of different frequency
components of the frequency domain signal Here we use
thescaled periodogram to calculate the intensity of the sig
nal in the frequency domain The formula is as follows
Pgωk 4
Fdωk2 k 0 F1 8
Then we compare Pgωkkand filter out the frequency
components ωkcorresponding to the first mf2 maxima
Their average value df
jPωkmfis noted as the defor
mation frequency of the trajectory We sort all the points
according to the magnitude of the deformation frequencies
df
jP
j1from low to high defining the points with low fre
quency of αr100 asnearly rigid points and denote theset of their subscripts as Ar Here αr01is a hyperpa
rameter that depends on the deformation properties
In this way the whole object is divided into a nearly
rigid region and a nonrigid region as the example shown
in Fig 2a We summarize the operation above as
ArAnrs
ˆS αr mf
9
333 Kernel Based Proxy Shape
After completing the segmentation we compute the proxy
shape SiofˆSi We model the weight matrix Λby the kernel
function kTiTj where Tjis the 3D trajectory of the jth
point The weight matrix can be expressed as follows
Λ
ϕT1 ϕT1 ϕT1 ϕTP

ϕTP ϕT1 ϕTP ϕTP
10
where ϕrepresents a feature map is the vector inner
product We use the inner product to compute the correla
tion between the vectors after the feature mapping as the
combination weights between different points in the space
In order to increase the difference between the lowrank
constraint effect on the nearly rigid and nonrigid regions
we set up the feature mapping ϕin the following form
ϕTi p
1δ2r1iδr1P1 iAr
δnr1P1 i Ar11
where 1iis a01vector in P 1 dimensions taking 1
only in the ith dimension δris a constant and δnris set
to1p
1αrP Here we provide a brief discussion of
model 10 When αr 1ie all points are in set ArSi
is just a shrink state of the original shape If αr1 all
points in the nonrigid region degenerate into the same su
per point as shown in Fig 2b The super point prevents
the lowrank constraint from acting directly on each point
in the nonrigid region which can help reduce the intensity
of the lowrank constraint on the nonrigid region In addi
tion the proxy shape calculation model 10 ensures that the
super point and nearly rigid points remain a unified whole2
34 Complete Model and Solution
In this section we integrate the complete model with the
unified framework 2 and perform the solution For the
data term F we use the reprojection constraint as the op
timization target The SWNN model 6 is set to the regu
larization term Gand constrains the shape sequence under
canonical coordinate The link between the camera coordi
nate and the canonical coordinate is constructed through the

the reshuffled sequence as ˆSgˆSRF3P Differ

ent from existing methods we propose a SpatialWeighted

Nuclear Norm SWNN optimization model that imposes

lowrank constraint on the reshuffle of the proxy shapes

SgˆSΛinstead of ˆS so the new lowrank optimization

model can be written as

min

ˆSS 6

The key to the model 6 is how to design the spatial weight

matrix Λ Next we present a method of constructing proxy

shapes Si i 1 Fto relax the lowrank constraint on

severe deformation regions We first introduce an algorithm

for segmenting objects according to the level of deformation

in different regions Based on this a kernel based method

for the construction of proxy shapes is proposed

332 Nearly Rigid Region Segmentation

In reality not all regions of nonrigid objects undergo the

same deformations but are spatiallyvariant For example

a talking mouth on a human face will undergo more severe

deformation than the nose Thus different regions should

fit the lowrank regularization differently In this section

we introduce the partition algorithm for different regions of

nonrigid objects see Fig 2a for the process

We define Tjˆ stjR31F

t1as the jth column of

shape sequence ˆS which represents a 3D trajectory The

xyz coordinates of a 3D trajectory can be viewed as time

domain signals and can be transformed into the frequency

domain by Discrete Fourier Transform DFT

dkF F12PF

t1ˆ stje2πitkF k 0 F17

we define ωkkF which symbolize the different fre

quency components Intuitively when the time domain sig

nal changes drastically its corresponding frequency domain

signal will contain more highfrequency components We

can compare the severity of different time domain signal

changes by calculating the intensity of different frequency

components of the frequency domain signal Here we use

thescaled periodogram to calculate the intensity of the sig

nal in the frequency domain The formula is as follows

Pgωk 4

Fdωk2 k 0 F1 8

Then we compare Pgωkkand filter out the frequency

components ωkcorresponding to the first mf2 maxima

Their average value df

jPωkmfis noted as the defor

mation frequency of the trajectory We sort all the points

according to the magnitude of the deformation frequencies

df

jP

j1from low to high defining the points with low fre

quency of αr100 asnearly rigid points and denote theset of their subscripts as Ar Here αr01is a hyperpa

rameter that depends on the deformation properties

In this way the whole object is divided into a nearly

rigid region and a nonrigid region as the example shown

in Fig 2a We summarize the operation above as

ArAnrs

ˆS αr mf

9

333 Kernel Based Proxy Shape

After completing the segmentation we compute the proxy

shape SiofˆSi We model the weight matrix Λby the kernel

function kTiTj where Tjis the 3D trajectory of the jth

point The weight matrix can be expressed as follows

Λ

ϕT1 ϕT1 ϕT1 ϕTP

ϕTP ϕT1 ϕTP ϕTP

10

where ϕrepresents a feature map is the vector inner

product We use the inner product to compute the correla

tion between the vectors after the feature mapping as the

combination weights between different points in the space

In order to increase the difference between the lowrank

constraint effect on the nearly rigid and nonrigid regions

we set up the feature mapping ϕin the following form

ϕTi p

1δ2r1iδr1P1 iAr

δnr1P1 i Ar11

where 1iis a01vector in P 1 dimensions taking 1

only in the ith dimension δris a constant and δnris set

to1p

1αrP Here we provide a brief discussion of

model 10 When αr 1ie all points are in set ArSi

is just a shrink state of the original shape If αr1 all

points in the nonrigid region degenerate into the same su

per point as shown in Fig 2b The super point prevents

the lowrank constraint from acting directly on each point

in the nonrigid region which can help reduce the intensity

of the lowrank constraint on the nonrigid region In addi

tion the proxy shape calculation model 10 ensures that the

super point and nearly rigid points remain a unified whole2

34 Complete Model and Solution

In this section we integrate the complete model with the

unified framework 2 and perform the solution For the

data term F we use the reprojection constraint as the op

timization target The SWNN model 6 is set to the regu

larization term Gand constrains the shape sequence under

canonical coordinate The link between the camera coordi

nate and the canonical coordinate is constructed through the