text
stringlengths
0
89.3k
that are more general and insensitive to the statistical prop
erties of the data is a feasible direction for improvement
9 Additional Experiments
In this section we first add the details of dataset processing
and algorithm implementation and then report additional
experiment results
91 Implementation Details
For parameters in ADMM Algorithm 2 we refer to 27 to
initialize β1e4βmax1e10Λ11and set Yn3
n1
to zero matrices The weights µ1 µ2 µ3of the optimization
objective 20 are set to 10101by default We can adjust
the sizes of µ2 µ3according to the shape sequences low
rank and smoothing properties In addition we found if
the shape sequence does not satisfy the lowrank property
well increasing the weight µ1of the reprojection term can
effectively improve the reconstruction results egµ11e1generally on the NRSfM Challenge dataset and µ11e2on
kinect and rug in the Semidense dataset For the estimation
of proxy shape we adjust the ratio of nearly rigid points αr
between 0to1depending on the deformation characteristics
of the object δris set to1
3on Mocap and NRSfM Challenge
datasets and 0on Semidense and H3WB datasets
As described in Sec 8 αrdetermines the optimization
approach used for different regions of the 3D structure so
αrshould depend on the specific characteristics of the ob
jects deformation eg spatial coherence continuity etc I
adopted a testing interval of 02for parameter selection ini
tially followed by a finer search around promising results
using a step size of 01 This granularity proved sufficient
for achieving good outcomes eliminating the need for fur
ther refinement in the search steps The stability test for our
method on αris illustrated in Fig 12 In most cases the al
gorithm is stable with respect to αronce a rough selection
interval has been determined and thus careful screening of
αris not necessary When we have a large amount of data
we can segment the keypoints by learning a certain distri
bution without setting hyperparameters like 13
92 H3WB Dataset Processing
The H3WB dataset 50 is an entire human body 3D dataset
extended from the H36M dataset 22 However this
dataset provides 2D annotations and their corresponding
3D structures frame by frame rather than in a sequence
We screened with the criterion of being as contiguous as
possible and obtained five sequences from S1 Eating2 S6
Smoking S6 Smoking1 S6 Directions and S7 Wait
ing2 Since the motion amplitude of these sequences is vast
we remove the frames with significant mutations and use
spline functions to interpolate the remaining parts to ob
tain five more realistic human action sequences Eating2
185 frames Smoking 180 Smoking1 265 Directions
245 and Waiting2 335 To obtain the coordinates of the
2D keypoints under the orthogonal projection model we set
up the camera motion in reference to 2
We show a partial 3D shape sequences of the H3WB
NRSfM dataset in Fig 7 We used the experimental setup
in Sec 62 to test the lowrank and smoothing properties of
the H3WB dataset and the results are displayed in Fig 13
The sequences in the H3WB dataset generally have higher
nuclear norm and firstorder smoothing errors compared to
the pickup sequence in Mocap Recovering the 3D shapes
in the H3WB dataset using lowrank and smoothing con
straints is more complicated However it is worth noting
that the TPAaligned sequences have better lowrank and
smoothing properties than the GT sequences which some
what guarantees the validity of the lowrank regularizationFigure 10 Qualitative Results on the NRSfM Challenge Benchmark The first row shows the images in the dataset and the second row
shows the reconstruction results of our method compared to the GT
Figure 11 Qualitative Results on the Semidense and H3WB dataset The reconstruction results of our method on the Semidense dataset
are closer to the GT shapes while the H3WB dataset is very challenging and cannot yet be accurately reconstructed
93 Additional Experiments on Missing Data
In reality the movement of an object often leads to occlu
sion of different regions resulting in missing 2D observa
tions obtained from camera shots Therefore the stability
of the algorithm on the missing dataset is significant We
show some tests on missing data in the main text and we
will add more quantitative results in this section
We use the model in Sec 72 and follow the settings
in 30 to randomly add masks occlusion rate 30 to the2D observation matrix W Before reconstruction we first
solve the lowrank approximation of the observation matrix
to complement it 9 which is important for the initializa
tion of the camera motion Tab 7 shows the test results
on the remaining sequences of the Mocap dataset and the
H3WB dataset Our method achieves the best performance
on sequences other than Shark and Walking The experi
mental results show that the reconstruction errors on data
with and without occlusion display consistency And the
comparison between them indicates that our method stillTable 7 Comparison of reconstruction errors on missing data in Mocap and H3WB datasets indicates that test results are
unavailable Oursw and Ourswo represent the test results with and without occlusion respectively
Method Shark Face Walking Eating2 Smoking1 Directions Smoking Waiting2
CSF217 00653 00412 01033 02556 03659 03025 04058 01862
PND30 00166 00177 00469 02167 02954 03054 04863 02256
PMP31 00116 00174 00507
Oursw 00268 00154 00796 01657 01971 02858 01804 01113
Ourswo 00258 00144 00710 01767 02040 02688 01693 01097
a Pickup 05 b Yoga 08
c Stretch 08 d Dance 02
e Articulf 01 f Stretchz 01
Figure 12 Stability test for hyperparameter αron different se
quences Straight lines indicate the results without the SWNN
module and broken lines indicate the reconstruction errors for dif