tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
dated Yn3
n1are Lagrange multipliers For the update
ofS we use the rectification algorithm in 26 The dif
ference is that when calculating the weights Θ we truncate
the singular values according to the shape basis dimension
Ksand normalize the weights In real scenes the captured
images are often obscured and it is difficult to observe all
the keypoints in each frame Our proposed framework can
handle the problem of missing points by simply adding vis
ible information to the data term More details and formulas
are provided in the Supp Sec 24 Experiments
41 Implementation Details and Evaluation Metric
Implementation Details The parameter settings in the
ADMM optimization algorithm are the same as in the Or
ganic Priors Method OPM 27 The model 12 is a
nonconvex optimization that requires the initialization of
camera motion and 3D shapes We use the camera mo
tion estimation algorithm in BMM 11 to initialize Rp
To build the weight matrix Λ a good initialization of the
shape sequence ˆSis needed to calculate the segmentation
of the nonrigid region Since our model is a unified frame
work there is no need to use other methods which can be
accomplished using only model 12 As shown in Algo
rithm 1 we first fix the correction rotation Qand set Λto
the Identity matrix After convergence the weight matrix
is calculated and all parameters Ωare well initialized In
addition βdin Algorithm 1 is generally 1e2or1e0 and
Ψ µ1 µ2 µ3 αr δr Ksis adjusted to the dataset see
Supp Sec 4 for more settings
Evaluation Metric We follow the setup in 27 using the
mean normalized 3D reconstruction error metric to evalu
ate the shape reconstruction results on the motion capture
benchmark MoCap semidense and H3WB dataset The
metric is defined as e3d1
FPF
i1

Sest
iSgt
i

F

Sgt
i

F
andSest
iSgt
idenote the estimated 3D shape and the corre
sponding groundtruth GT value respectively We remove
the global ambiguity 2 23 as in 27 before computing
the 3D reconstruction error To evaluate our approach on
the NRSfM benchmark dataset 23 we use the officially
supplied metric script
42 Datasets and Results
MoCap Benchmark Dataset This dataset is a standard
benchmark for NRSfM consisting of 8 real sequences
Akhter et al 2 introduced five sequences Drink Pickup
Yoga Stretch and Dance And the other three Face Walk
ing and Shark were presented by Torresani et al 45
Tab 1 and Fig 3a demonstrate the reconstruction errors e3d
of our method compared to other methods and some visual
results respectively As shown in Tab 1 our method per
forms best or secondbest across multiple sequences indi
cating that our method is able to accommodate diverse types
of deformation Our method also achieves comparable re
sults in sequences such as Shark and Walking outperform
ing the pure lowrank constraint methods 11 26 27
NRSfM Challenge Dataset Jensen et al 23 recently
proposed a new challenging benchmark This dataset con
tains five types of nonrigid deformation Articulated Bal
loon Paper Stretch and Tearing Each subject contains six
observation sequences captured by different types of cam
era motion ie circle flyby line semicircle tricky and
zigzag For each subject we calculate the reconstructionTable 1 3D reconstruction errors on MoCap dataset Our method shows advantages over many matrix factorization methods and
Procrustean alignment methods The secondbest results are underlined and the shape basis dimension Ksis shown in brackets
Data CSF1 15 CSF2 17 KSTA 16 PND 30 PMP 31 CNS 32 PR 41 BMM 11 RBMM 26 OPM 27 Ours
Drink 00223 00223 00156 00037 00018 00431 00063 00152 00119 00071 00031 13
Pickup 02301 02277 02322 00372 00127 01281 00157 00315 00198 00152 00126 12
Yoga 01467 01464 01476 00140 00128 01845 00175 00225 00129 00122 00109 10
Stretch 00710 00685 00674 00156 00124 00939 00156 00247 00144 00124 00114 12
Dance 02705 01983 02504 01454 01278 00759 01266 01445 01491 001209 00921 13
Face 00363 00314 00339 00165 00166 00248 00164 00206 00179 00145 00144 5
Walking 01893 01035 01029 00465 00424 00396 00544 00908 00882 00816 007104
Shark 00081 00444 00160 00135 00099 00832 00272 02311 00551 00550 002586
Table 2 Reconstruction error comparison with stateoftheart on
NRSfM Challenge dataset We report the results in millimeters
Data CSF2 17 BMM 11 RBMM 26 AOW 21 BP 35 OPM 27 Ours
Articul 1152 1849 1600 1503 1610 1218 1069
Balloon 1014 1039 784 805 829 629 728
Paper 972 894 1069 1045 670 886 791
Stretch 865 1002 753 901 766 636 543
Tearing 1204 1423 1634 1620 1126 1091 1077
Table 3 Mean normalized 3D reconstruction errors on Semidense
dataset indicates the estimation failed due to excessive com
putational overhead
Data CSF2 17 BMM 11 CNS 32 RBMM 26 Ours OursI
Kinect 00232 01212 00453 00199 00356 00161
Rug 00189 00109 00135 00088 00088

dated Yn3

n1are Lagrange multipliers For the update

ofS we use the rectification algorithm in 26 The dif

ference is that when calculating the weights Θ we truncate

the singular values according to the shape basis dimension

Ksand normalize the weights In real scenes the captured

images are often obscured and it is difficult to observe all

the keypoints in each frame Our proposed framework can

handle the problem of missing points by simply adding vis

ible information to the data term More details and formulas

are provided in the Supp Sec 24 Experiments

41 Implementation Details and Evaluation Metric

Implementation Details The parameter settings in the

ADMM optimization algorithm are the same as in the Or

ganic Priors Method OPM 27 The model 12 is a

nonconvex optimization that requires the initialization of

camera motion and 3D shapes We use the camera mo

tion estimation algorithm in BMM 11 to initialize Rp

To build the weight matrix Λ a good initialization of the

shape sequence ˆSis needed to calculate the segmentation

of the nonrigid region Since our model is a unified frame

work there is no need to use other methods which can be

accomplished using only model 12 As shown in Algo

rithm 1 we first fix the correction rotation Qand set Λto

the Identity matrix After convergence the weight matrix

is calculated and all parameters Ωare well initialized In

addition βdin Algorithm 1 is generally 1e2or1e0 and

Ψ µ1 µ2 µ3 αr δr Ksis adjusted to the dataset see

Supp Sec 4 for more settings

Evaluation Metric We follow the setup in 27 using the

mean normalized 3D reconstruction error metric to evalu

ate the shape reconstruction results on the motion capture

benchmark MoCap semidense and H3WB dataset The

metric is defined as e3d1

FPF

i1

Sest

iSgt

i

F

Sgt

i

F

andSest

iSgt

idenote the estimated 3D shape and the corre

sponding groundtruth GT value respectively We remove

the global ambiguity 2 23 as in 27 before computing

the 3D reconstruction error To evaluate our approach on

the NRSfM benchmark dataset 23 we use the officially

supplied metric script

42 Datasets and Results

MoCap Benchmark Dataset This dataset is a standard

benchmark for NRSfM consisting of 8 real sequences

Akhter et al 2 introduced five sequences Drink Pickup

Yoga Stretch and Dance And the other three Face Walk

ing and Shark were presented by Torresani et al 45

Tab 1 and Fig 3a demonstrate the reconstruction errors e3d

of our method compared to other methods and some visual

results respectively As shown in Tab 1 our method per

forms best or secondbest across multiple sequences indi

cating that our method is able to accommodate diverse types

of deformation Our method also achieves comparable re

sults in sequences such as Shark and Walking outperform

ing the pure lowrank constraint methods 11 26 27

NRSfM Challenge Dataset Jensen et al 23 recently

proposed a new challenging benchmark This dataset con

tains five types of nonrigid deformation Articulated Bal

loon Paper Stretch and Tearing Each subject contains six

observation sequences captured by different types of cam

era motion ie circle flyby line semicircle tricky and

zigzag For each subject we calculate the reconstructionTable 1 3D reconstruction errors on MoCap dataset Our method shows advantages over many matrix factorization methods and

Procrustean alignment methods The secondbest results are underlined and the shape basis dimension Ksis shown in brackets

Data CSF1 15 CSF2 17 KSTA 16 PND 30 PMP 31 CNS 32 PR 41 BMM 11 RBMM 26 OPM 27 Ours

Drink 00223 00223 00156 00037 00018 00431 00063 00152 00119 00071 00031 13

Pickup 02301 02277 02322 00372 00127 01281 00157 00315 00198 00152 00126 12

Yoga 01467 01464 01476 00140 00128 01845 00175 00225 00129 00122 00109 10

Stretch 00710 00685 00674 00156 00124 00939 00156 00247 00144 00124 00114 12

Dance 02705 01983 02504 01454 01278 00759 01266 01445 01491 001209 00921 13

Face 00363 00314 00339 00165 00166 00248 00164 00206 00179 00145 00144 5

Walking 01893 01035 01029 00465 00424 00396 00544 00908 00882 00816 007104

Shark 00081 00444 00160 00135 00099 00832 00272 02311 00551 00550 002586

Table 2 Reconstruction error comparison with stateoftheart on

NRSfM Challenge dataset We report the results in millimeters

Data CSF2 17 BMM 11 RBMM 26 AOW 21 BP 35 OPM 27 Ours

Articul 1152 1849 1600 1503 1610 1218 1069

Balloon 1014 1039 784 805 829 629 728

Paper 972 894 1069 1045 670 886 791

Stretch 865 1002 753 901 766 636 543

Tearing 1204 1423 1634 1620 1126 1091 1077

Table 3 Mean normalized 3D reconstruction errors on Semidense

dataset indicates the estimation failed due to excessive com

putational overhead

Data CSF2 17 BMM 11 CNS 32 RBMM 26 Ours OursI

Kinect 00232 01212 00453 00199 00356 00161

Rug 00189 00109 00135 00088 00088