Datasets:

tensorkelechi

/

arxiv-cstext

like 0

Proxy ShapeSuper Point

iSEnhance

Connection

Degenerate

Nearly Rigid Point b Construction for proxy Shape

Figure 2 Overview of region segmentation and proxy shape construction a We use DFT to analyze the 3D trajectories dividing them

by comparing the frequency components contained in each trajectory The figure shows a segmentation result b The geometric center of

the nonrigid region is set as the super point and it is linearly combined with nearly rigid points to construct the proxy shape

other deformation regularization constraints should be sat

isfied in the smooth motion sequence We propose a cam

era motion refinement module called Temporallysmooth

Procrustean Alignment TPA which aligns the shape se

quence under the smoothness constraint as follows

LtpaRS 1

2F1X

i1RiSiRi1Si12

F3

where RdiagRiRiSO3 The TPA module

3 estimates alignment matrix Rby the principle of mini

mizing the difference between the 3D shapes of consecutive

frames after rotation transformation as in Fig 1 b Dif

ferent from GPA 1 TPA does not need to estimate the

mean shape of the sequence which greatly simplifies the

optimization and avoids overpenalizing nonisotropic de

formations In addition we have experimentally verified

that sequences aligned by the TPA module can have more

similar lowrank and smoothing properties to the real se

quences compared with GPA see Sec 43 and Fig 4

Under the NRSfM problem setting the TPA module can

gradually put the shape sequence in a smooth state by ad

justing the camera motion Assume that the estimated cam

era poses are Ri i 1 Fand define the initialization

of the alignment transformation as RpiRT

i We denote

RpiSiasSi where Siis in the camera coordinate Then

we use the TPA module to estimate the correction rotation

QdiagQias follows

LQ

tpa

QS

1

2F1X

i1

QiSiQi1Si1

2

F

1

2F1X

i1PX

j1Qi sijQi1 si1j2

24

where sijR31is the jth column of Si We take the Lie

algebra ϕiso3ofQias the optimization variable andthe derivative of the optimization objective LQ

tpa4 with

respect to ϕican be derived as

LQ

tpa

ϕiPX

j1

Qi sijT

r1

ijr0

ij

5

where rm

ijQim simjQim1 sim1j and a

represents the skewsymmetric matrix of vector a Once the

gradient is calculated we can use traditional numerical op

timization algorithms to solve model 4 eg BFGS LMF

etc 33 see more derivation details and discussions in the

Supplementary Material Supp Sec 1

TPA largely separates the rigid motion between consec

utive frames in the sequence Sand makes it gradually con

verge to a smooth state After reducing the interference of

rigid motion we can better utilize the lowrank constraint to

recover nonrigid deformation of the aligned sequence QS

In the following we introduce the shape estimation method

and unify them afterward

33 SpatialWeighted Nonrigid Structure Model

331 SpatialWeighted Nuclear Norm

In realworld scenarios lowrank regularization can han

dle many nonrigid deformation situations but always over

penalize drastic deformations In the absence of priors to

reconstruct the deformation we can give higher degrees of

freedom for shape recovery by relaxing the lowrank con

straint An object whose overall or local deformation is very

drastic can lead to its 3D shape sequence ˆSin the world co

ordinate not satisfying the lowrank constraint But we can

transform its spatial structure by linearly combining the tra

jectories of ˆSto degenerate the object to a lowrank state

Given the affinity matrix ΛRPP we define the sequence

ofproxy shape asSˆSΛ We can then utilize lowrank

constraint to recover the proxy shapes which in turn re

duces the strength of the constraint in the region of severe

deformation to better recover ˆSWe use the nuclear norm with the rearrangement op

erator g11 to build the lowrank constraint and denote