text
stringlengths
0
89.3k
Mat 00620 01088 00197 00285 00238 00182
Table 4 Comparison of mean normalized 3D reconstruction error
with wellknown sparse NRSfM methods on H3WB dataset
Data CSF2 17 BMM 11 PND 30 CNS 32 RBMM 26 Ours
Eating2 02662 02084 02309 02250 01779 01767
Smoking1 03534 02287 03397 02982 02290 02040
Directions 02953 02872 03002 02674 02964 02688
Smoking 01980 01759 04349 01968 01760 01693
Waiting2 02125 01712 03306 01848 01478 01097
Mean 02651 02143 03272 02344 02054 01857
errors under the six camera motions and take the average
as the final error for that subject The quantitative compari
son with other methods and the qualitative visualization are
shown in Tab 2 and Fig 3b Our method achieves the best
results on Articul Stretch and Tearing and also has high
reconstruction accuracy on Balloon and Paper The numer
ical results further show the advantages of our method on
multicategory deformation reconstruction
Semidense Dataset We evaluated our method on the
kinect paper rug table mat datasets 38 19115960
frames and 150339121500 keypoints respectively We
compared our method with other wellknown sparse
NRSfM methods PND 30 fails on this dataset due to ex
cessive computational overhead caused by the large number
of points see the results in Tab 3 and Fig 3c We denote
our method that initializes Rpusing the modified algorithm
in 26 as OurI Our proposed model is also effective in re
ducing the reconstruction error on semidense dataset Our
method can significantly reduce the effect of poor camera
motion on shape estimation and also further improve recon
struction accuracy with better motion initialization
a Visualization on MoCap Dataset
b Visualization on NRSfM Challenge Dataset
c Visualization on Semidense and H3WB Dataset
Figure 3 a Visualization of Shark and Dance results b Visu
alization of Balloon and Articulated results Top row is the image
in dataset and bottom row is the 3D reconstruction shape c Vi
sualization of Rug and Eating2 results
H3WB Dataset Human36M 3D WholeBody H3WB
dataset 50 is an entire human body 3D dataset includ
ing face hands body and feet H3WB is an extension of
Human36M 22 and contains multiple categories of com
mon life actions such as Directions Eating and Smoking
The human body in H3WB is annotated with 133keypoints
17for body 6for feet 68for face and 42for hands We
selected five sequences in H3WB and processed them to
obtain a new evaluation dataset Eating2 Smoking1 Di
rections Smoking Waiting2 185265245180 and335
frames respectively see Supp Sec 4 for more detailsTable 5 Comparison of reconstruction errors with and without SWNN module The table shows the comparison results on the MoCap
dataset and the Articulcircle Articulflyby Stretchzigzag sequences in NRSfM Challenge dataset
Method Drink Pickup Yoga Stretch Dance Face Articulc Articulf Stretchz
Oursw 00031 00126 00109 00114 00921 00144 23442 38229 06460
Ourswo 00050 00137 00120 00132 00932 00161 35626 48042 07341
Figure 4 Shape alignment test for the TPA module We compare
the lowrank and smoothing properties of GT nonrigid sequences
randomly rotated disrupted sequences and sequences aligned by
TPA or GPA The results show that the TPAaligned sequences
have more similar properties to the GT sequences
As manifested in Tab 4 our method shows better recon
struction results than other priorfree and procrusteanbased
methods Visualization illustrated in Fig 3c
43 Ablation Study
Role of TPA Module First we verify the alignment func
tion of TPA module 4 on Pickup and Yoga sequences We
randomly sample Lie algebra pnR3Ffrom Gaussian
noiseN001 and then map it to a 3D rotation matrix
RdR3F3Fusing Rodrigues rotation formula We de
fine the shape sequence disrupted by random rotations as
Snoise RdSgt where SgtR3FPis the GT sequence
in the world coordinate We realigned the sequence with the
TPA module and compared the result with GPA 30 Fig 4
compares the nuclear norm iegand the firstorder
smoothness 11 of SgtSnoise and the sequences aligned
by the two algorithms showing that the TPAaligned se
quence has the most similar properties to the GT sequence
We also test the stability of the TPA module for cam
era motion initialization We used the same approach to
sample perturbed rotations Rdwith standard deviation σ
of 01 02 and 05 Adding ambiguity Rdto the camera
motion initialized by BMM 11 we then compare the four
methods BMM RBMM BMMSmooth Add firstorder
smoothness constraint BMMTPA For each σ we exper
imented five times and calculated the mean the results are
shown in Tab 6 Compared to other methods the TPA mod
ule maintains its effectiveness even when a large deviation
has been applied to the camera motion This shows that our
method can effectively deal with the rotation ambiguity in
camera motion estimation and can improve the effective
ness and stability of the smoothing constraint
Effectiveness of SWNN We compared the two cases of usTable 6 Experiment for the effectiveness of the TPA module in
correcting camera motion Adjustment for weight µ3in brackets
SD BMM 11 RBMM 26 BMMSmooth BMMTPA
σ 01 00493 00507 00845 00237 1e1
σ 02 00845 00769 01149 00592 1e2
σ 05 01743 01681 01905 01665 1e3
ae3dwith noise
be3dwith occlusion
Figure 5 a Shape errors on noisy sequences b Experiment for
3D reconstruction on missing data
ing and not using with ΛI the SWNN module in model