Human Pose Estimation through a Novel Multi-view Scheme

Jorge L. Charco

1,2

, Angel D. Sappa

1,3

and Boris X. Vintimilla

ESPOL Polytechnic University, Escuela Superior Polit

ecnica del Litoral, ESPOL,

Campus Gustavo Galindo Km. 30.5 V

ıa Perimetral, P.O. Box 09-01-5863, Guayaquil, Ecuador

Universidad de Guayaquil, Delta and Kennedy Av., P.B. EC090514, Guayaquil, Ecuador

Computer Vision Center, Ediﬁci O, Campus UAB, 08193 Bellaterra, Barcelona, Spain

Keywords:

Multi-view Scheme, Human Pose Estimation, Relative Camera Pose, Monocular Approach.

Abstract:

This paper presents a multi-view scheme to tackle the challenging problem of the self-occlusion in human

pose estimation problem. The proposed approach ﬁrst obtains the human body joints of a set of images,

which are captured from different views at the same time. Then, it enhances the obtained joints by using a

multi-view scheme. Basically, the joints from a given view are used to enhance poorly estimated joints from

another view, especially intended to tackle the self occlusions cases. A network architecture initially proposed

for the monocular case is adapted to be used in the proposed multi-view scheme. Experimental results and

comparisons with the state-of-the-art approaches on Human3.6m dataset are presented showing improvements

in the accuracy of body joints estimations.

1 INTRODUCTION

The 2D Human Pose Estimation (HPE) problem is

generally tackled by ﬁrst detecting human body joints

(e.g., wrist, shoulder, knee, etc.) and then connect-

ing them to build the human body stick ﬁgure. Differ-

ent solutions have been proposed in the literature (e.g.

OpenPose, DeepPose, Stacked Hourglass Networks)

and robust solutions obtained when all body joints are

detected. However, this problem becomes a challeng-

ing one when joints are occluded (e.g., due to self oc-

clusions, which is something common in monocular

vision system scenarios). Applications such as human

action recognition, augmented reality, healthcare, just

to mention a few, have taken advantage of the accu-

racy of 2D human pose to develop on top of them dif-

ferent solutions. In recent years, convolutional neural

networks (CNN) have become a de facto tool to tackle

most of computer vision tasks; for instance it has been

used in image enhancement, object detection, camera

pose estimation, just to mention a few, getting better

results with respect to classical approaches (e.g.,(Tian

et al., 2019), (Wu et al., 2020), (Charco et al., 2018)).

Also in the human pose estimation problem we can

ﬁnd different CNN architectures to solve it in a deep

learning based framework showing appealing results

(e.g., (Wei et al., 2016), (Newell et al., 2016), (Fang

et al., 2017), (Cao et al., 2019), (Sun et al., 2019)).

The proposed approaches have used as input a set

of images with single or multiple-person to feed the

architectures, typically from single-view. Regarding

this latter point, multiple-person pose estimation, the

number of people in the image increase the compu-

tational cost, and hence, also the inference time in

real-time. In order to tackle these problems, two ap-

proaches have been introduced. The ﬁrst, known as

top-down, localizes the persons in the image and es-

timate the body joints. The second, referred to as

bottom-up, estimates the human body parts in the im-

age and then compute the pose. Despite appealing

results obtained on HPE from single-view, the chal-

lenge lies in the occlusions of the human body joints

in complex poses, causing self-occlusions of certain

parts of the human body, in spite of the fact that also

the scene could contain multiples moving objects (i.e.,

bicycles, cars), leading partial occlusion of the human

body. In order to overcome this problem, multi-view

approaches could be considered; in these cases the

human body is captured at the same time from dif-

ferent positions by different cameras. Hence, joints

self-occluded in one view can be observed without

occlusion by some other camera from other point of

view.

The multi-view framework has been already ex-

plored to tackle the region occlusion problem in tasks

such as 3D-reconstruction, camera pose, autonomous

Charco, J., Sappa, A. and Vintimilla, B.

Human Pose Estimation through a Novel Multi-view Scheme.

DOI: 10.5220/0010899900003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 5: VISAPP, pages

855-862

ISBN: 978-989-758-555-5; ISSN: 2184-4321

855

driving, object detection. (e.g., (Xie et al., 2019),

(Sarmadi et al., 2019), (Charco et al., 2021), (Hof-

bauer et al., 2020), (Tang et al., 2018)). For the 2D hu-

man pose estimation problem by using a multi-view

approach, few works have been proposed. The au-

thors in (Qiu et al., 2019) have proposed a CNN ar-

chitecture to fuse all features on epipolar line of the

images across of all different views. In (He et al.,

2020), the authors have proposed to leverage the us-

age of the intermediate layer to ﬁnd its corresponding

point in a neighboring view, an then combine the fea-

tures of both views.

On the contrary to previous approaches, where

complex deep learning based architectures feed with

images from different cameras acquired at the same

time, in the current work a compact architecture, orig-

inally proposed for monocular scenarios, is adapted to

the multi-view scenario. Actually, the multi-view sce-

nario is considered just during the training stage. Im-

ages of the same scene, simultaneously acquired by

cameras at different point of views, are acquired and

used for the CNN training. The proposed architecture

uses a variant of ResNet-152 with learning weights as

backbone that was proposed in (Iskakov et al., 2019).

The multi-view adapted backbone proposed in the

current work is trained considering the set of images

acquired as mentioned above. It allows to tackle com-

plex poses and overcome the self-occlusion problem,

improving the accuracy of estimated joints, being the

basis to solve other related problems, such as 3D hu-

man pose estimation.

The remainder of the paper is organized as fol-

lows. In Section 2 previous works are summarized;

then, in Section 3 the proposed approach is detailed

together with a description of the scheme multi-view.

Experimental results are summarized in Section 4

together with comparisons with state-of-the-art ap-

proachs. Finally, conclusions and future work are

given in Section 5.

2 RELATED WORK

Vision-based human pose estimation is a challeng-

ing problem due to the complexity to extract fea-

tures from images; this complexity is due to differ-

ent lighting conditions, complex poses, occlusions,

among others. On this basis, CNN models have been

used for this purpose due to the capability of analysis

of images to extract key features of the human body

(e.g., joints) improving state-of-art results. Some

works have been proposed for 2D-human pose esti-

mation from a single-view scenario. The authors in

(Toshev and Szegedy, 2014) have proposed a Deep

Neural Network (DNN) as a regress to get the (x, y)

image coordinates of human body joints. Addition-

ally, they propose to use a scheme of a cascade of

DNN to increase the precision of estimated coordi-

nates by using higher resolution sub-images for re-

ﬁning the predicted joints. In (Tompson et al., 2015)

the authors have proposed a multi-resolution ConvNet

architecture to implement a sliding window detec-

tor with overlapping contexts to generate heatmaps

for each joint. The architecture is fed with images,

which are running through multiple resolution banks

in parallel, and thus capturing important features at

a variety of scales. The proposal is trained by min-

imizing the Mean Squared-Error (MSE) distance of

the predicted heatmap to a targeted heatmap. Simi-

larly to the previous works, the authors in (Carreira

et al., 2016) have proposed a convolutional network

that takes advantage of hierarchical feature extractor,

which introduces a top-down feedback of both input

and output spaces. The proposal estimates the current

human pose, and the joints with wrong predictions are

iteratively improved by feeding back error predictions

instead of trying to directly predict the target outputs.

The authors in (Newell et al., 2016) have proposed a

model that consists of steps of pooling and upsam-

pling layers, which are stacked together. The pro-

posed model extracts features at every scale to cap-

ture global and local information of the images. Skip

connections are used to preserve spatial information

at each resolution.

On the contrary to the previous approaches, in

(Xiao et al., 2018) an architecture have been pro-

posed, which consists of a variant of ResNet that in-

cludes a few deconvolutional layers at the end. These

simple changes preserve better the information for

each resolution than one with skip connections. Sim-

ilarly to the previous approaches, MSE is used as the

loss between the predicted heatmaps and the targeted

heatmaps. The authors in (Sun et al., 2019) have pro-

posed a novel architecture, which is able to maintain a

high-resolution representation through the whole pro-

cess, i.e., it starts from a high-resolution subnetwork

as the ﬁrst stage, and gradually add high-to-low reso-

lution subnetworks to form more stages that are con-

nected in parallel, instead of recovering the resolution

through a low-to-high process.

Just few works have been proposed to solve the

human pose estimation problem leveraging any ge-

ometry information of the cameras to improve the

2D detector. In (Qiu et al., 2019), the authors have

introduced a cross-view fusion scheme into CNN to

jointly estimate 2D poses from multiple views. The

initial pose heatmaps are generated for each image

into a multi-view scheme, the corresponding features

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

856

Figure 1: CNN backbone feeds with a set of pairs of images of the same scene simultaneously acquired from different points

of view. The multi-view fusion scheme allow to estimate occluded joints with information from other views across of the

relative camera pose.

between them are found on the epipolar line, and then

they are fused across different views. Similarly to the

previous work, the authors in (He et al., 2020) have

proposed an architecture to leverage 3D-aware feature

in the intermediate layers of the 2D detector, and not

only during the ﬁnal robust triangulation phase. The

epipolar transformer is used to augment the interme-

diate features of a 2D detector for a given view (ref-

erence view) with features from neighboring views

(source view). The authors in (Remelli et al., 2020)

have proposed a novel multi-camera fusion technique,

which uses the feature transform layers to map images

from multiple-views and exploit 3D geometry infor-

mation to a common canonical representation by ex-

plicitly conditioning them on the camera projection

matrix.

3 PROPOSED APPROACH

The proposed approach consists to leverage the multi-

view scheme to solve the self-occlusion problems in

the 2D human pose estimation. The CNN backbone,

proposed by (Iskakov et al., 2019), is used as baseline

to be retrained with the proposed multi-view scheme.

Basically, this backbone is a variant of Resnet-152

with learnable weights, transposed convolutions, and

the number of human body joints as the output chan-

nels. For more details see the work mentioned above.

A multi-view system of C calibrated and synchro-

nized cameras with known parameters R

(i.e., intrin-

sic and extrinsic parameters), which capture the per-

formance of a single-person in the scene from differ-

ent views, is used by the proposed model. The im-

ages acquired by the multi-view system are denoted

as Im

, and organized in pairs of images, which be-

long to different views, namely, reference view Im

re f

and source view Im

src

. The output of the backbone

is a set of heatmaps for each image. These heatmaps

correspond to each human body joint, which are fused

Figure 2: An image point p

src

back-projects to a ray in 3D

deﬁned by the point p

src

and depth (Z). The ray is projected

to the image plane of reference view to generate the epipolar

line (L).

across source view considering the conﬁdence of each

joint, doing robust the human pose of each view (see

Fig. 1). The details of proposed multi-view approach

are given below.

3.1 Multi-view Scheme

Given a set of pairs of images, the CNN

backbone extracts the heatmaps of each joint

for each input image separately, which are de-

noted as M

re f

, ..., Im

re f

and M

src

{

src

, ..., Im

src

}

, where i is the number of joints and,

re f and src correspond to the reference and source

view respectively. The heatmaps are used to estimate

the 2D positions of each joint for each input image.

First, it is computed the softmax across the spatial

axes; and then, the 2D positions of the joints (p

(x,y)

)

are computed as the center of mass of the correspond-

ing heatmaps, which is deﬁned as:

(x,y)

∑

u=1

∑

v=1

(u,v)

.(ζ

(

u,v)

)), (1)

Human Pose Estimation through a Novel Multi-view Scheme

857

where ζ

represents the function softmax; h repre-

sents the ROI of the heatmaps of i-th joint and W and

H correspond to the size of the heatmap ROI.

The values of the joints in the world coordinate

system P = (X, Y, Z) are obtained using each 2D po-

sition of each joint of each image, as show in Eq. (2):

= f

, (2)

where x, y are the 2D position of i-th joint obtained in

Eq. (1), and f corresponds to the focal length of the

camera. Since, the depth (Z) of the joint is unknown,

then two values are used to solve the Eq. (2). The

ﬁrst, a value of depth near zero that corresponds to

the position close of camera in the world coordinate

system; and the second, a value of depth near to the

size of the space of the scene. Empirically, this value

has been set to 10m for the experiments.

For each 2D position of the joint, two points in

the world coordinate system using the depth estima-

tion mentioned above are computed with Eq. (2),

then they are transformed by using the relative camera

pose between both views (reference and source view)

and projected to the image plane, as shown below:

rel

= Rot

src

· (T

re f

− T

src

), (3)

Rot

rel

= Q(Rot

re f

.T )

−1

∗ Q(Rot

src

.T ), (4)

re f

src

(x,y)

= ∆2D

re f

(Rot

rel

· (P

− T

rel

)), (5)

where Q(.) represents the quaternion. Rot ∈ R

3x3

and T ∈ R

3x1

represent the matrix rotation and the

vector translation respectively. P

corresponds to co-

ordinates of the i-th joint, obtained in Eq. (2), in the

world coordinate system. As the depth (Z) of i-th

joint is unknown, the linear equation on image plane

of reference view is calculated using the points ob-

tained in the Eq. (5). By deﬁnition, the depth (Z)

of the i-th joint in the source view should be any 2D-

point on the linear equation of reference view. Given

that the linear equation has inﬁnite points, and any

of them could be the depth of i-th joint in the source

view, then, the 2D-point on linear equation used as

the depth of the joint in the source view is calculated

by using the intersection between the 2D-point of the

joint calculated in the reference view p

re f

(x,y)

and the

linear equation previously obtained.

Since the i-th joint has two different 2D positions

in the image plane, the ﬁrst corresponds to the refer-

ence view p

re f

(x,y)

, and the second corresponds to the

source view, which are projected to the reference view

re f

src

(x,y)

by using Eq. (5), the conﬁdence values of i-

th are obtained (see Fig. 2). It is calculated as the

distance between the ground-truth of 2D position of

i-th joint and the 2D position of i-th joint obtained by

Eq. (1). In order to improve 2D position of i-th joint

in the reference view, the conﬁdence values and 2D

positions of i-th joints (p

re f

(x,y)

, p

re f

src

(x,y)

) are used, as

shown in Eq. (7).

ω = 1 −



∆

(

, γ

)

∑

∆

(

, γ

)



, (6)

upd

(x,y)

= ω ∗ p

(x,y)

, (7)

where (

γ , γ) represent the ground truth and prediction

of 2D position of i-th joint respectively, and ω cor-

responds to the conﬁdence of the points of i-th joint

in the reference view, including the points projected

from source view.

Note that δ

upd

(x,y)

corresponds to the new 2D po-

sitions of i-th joint, which has been enhanced with the

information and conﬁdence of i-th joint obtained from

the source view. Finally, the loss function used in the

proposed approach is deﬁned as:

Loss =

∑

i=1



upd

(x,y)

− ˆp

(x,y)



, (8)

where N corresponds to the number of joints, and

ˆp

(x,y)

is the ground-truth of i-th joint in image plane.

3.2 Dataset and Metrics

The experiments are conducted on one large-scale

pose estimation public dataset with multi-view syn-

chronized images and evaluated using the JDR(%)

metric. This section will breaﬂy describe both of

them, dataset and used metrics.

3.2.1 Human 3.6m

The Human3.6m dataset was proposed by (Ionescu

et al., 2014), and it is currently one of the largest

publicly available human pose estimation benchmark.

It can be used with monocular or multi-view se-

tups. Four synchronized and calibrated digital cam-

eras were used to capture 3.6 million frames with a

single-person. The motions are performed by 11 pro-

fessional actors (6 males, 5 female) in different activi-

ties such as taking photo, discussion, smoking among

other. In the current work, subjects 1, 5, 7, and 8 are

used for trained the proposed approach; while sub-

jects 9 and 11 are used just for testing. Images from

all the cameras are used during the training and test-

ing process.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

858

Table 1: Comparison of 2D pose estimation accuracy on Human3.6m dataset using JDR(%) as metric. ” − ”: these entries

were absent. ∗: approach presented in (Qiu et al., 2019). ϒ trained again by (He et al., 2020). ψ approach presented in (He

et al., 2020). R50 and R152 are ResNet-50 and ResNet-152 respectively. Scale is the input resolution of the network.

Net scale shlder elb wri hip knee ankle root neck head Avg

Sum epipolar line

∗

R152 320 91.36 91.23 89.63 96.19 94.14 90.38 - - - -

Max epipolar line

∗

R152 320 92.67 92.45 91.57 97.69 95.01 91.88 - - - -

Cross-View fusion

∗ϒ

R50 320 95.6 95.0 93.7 96.6 95.5 92.8 96.7 96.5 96.2 95.9

Cross-View fusion

∗ϒ

R50 256 86.1 86.5 82.4 96.7 91.5 79.0 100 93.7 95.5 95.1

Epipolar transformer

R50 256 96.44 94.16 92.16 98.95 97.26 96.62 99.89 99.68 99.63 97.01

Mview-Joints (ours) R152 384 99.65 97.31 93.70 99.22 97.24 97.45 99.83 99.82 99.75 98.22

Table 2: Comparison of average median Euclidean distance error between Mview-Joints and Learning triangulation backbone

proposed by (Iskakov et al., 2019) on Human3.6m. Backbone: Resnet 152 with pretrained weight (Iskakov et al., 2019).

Net shlder elb wri hip knee ankle root neck nose belly head Avg

Learning

triangulation

Backbone 7.84 8.00 7.40 7.55 7.45 9.70 5.75 5.86 6.46 6.47 6.57 7.18

Mview-Joints

(ours)

Backbone +

Multi-view

7.88 6.73 7.08 7.62 6.82 9.19 5.24 6.05 5.29 6.15 3.25 6.48

3.2.2 Metrics

The metric to be used to evaluate the performance

of the obtained results is an important factor. In the

human pose estimation problem, the Joint Detection

Rate (JDR) is generally used. The JDR measures the

percentage of success f ully detected joints, assuming

as a successful detection those joints where the dis-

tance between the estimated and the ground truth joint

is smaller than a given threshold; in the current work

this threshold has been deﬁned as half of the head

size, as proposed in [2]. In the current work, in ad-

dition to the JDR, the Euclidean distance error for

every estimated joint with respect to the correspond-

ing ground truth has been also computed. These Eu-

clidean distance error values help to determine the ac-

curacy of each joint of the estimated human pose.

4 EXPERIMENTAL RESULTS

As mentioned above, the multi-view approach is pro-

posed to tackle challenging scenarios where self-

occlusions of joints happen resulting in difﬁcult 2D

human pose estimation. This section presents de-

tails on the experimental results by training the pro-

posed multi-view scheme with Human 3.6m dataset

(Ionescu et al., 2014). The proposed approach was

implemented with Pytorch and trained with NVIDIA

Titan XP GPU and Intel Core I9 3.3GHz CPU. Adam

optimizer is used to train the network with a learn-

ing rate of 10

−5

and batch size of 32 (i.e., eight hu-

man poses simultaneously captured from four differ-

ent points of view).

4.1 Training of Multi-view Scheme

The CNN backbone used in the proposed architecture

was initialized with the weights of Resnet-152 pre-

trained by (Iskakov et al., 2019). The network ar-

chitecture was trained on Human 3.6m dataset. As

pre-processing dataset, the images were cropped ac-

cording to the bounding box of the person and resized

to 384x384 pixels; then, the mean value of intensity

of pixels was computed and subtracted from the im-

ages. For the training process, a set of 60k images

were used to feed to the network, which was trained

until 20 epochs; it takes about 120 hours. The pre-

processing mentioned above has been also used dur-

ing the evaluation phase. In the evaluation a set of 8k

images have been considered.

4.2 Results and Comparisons

Experimental results obtained with the proposed ar-

chitecture are presented in Table 1, which shows some

joints and compare them with state-of-the-art CNN-

based methods by using the JDR metric. The pro-

posed approach referred to as Mview-Joints outper-

forms the previous works on most of body joints. The

improvement is most signiﬁcant for the shoulder, el-

bow, and ankle joints, which increment from 96.44%

to 99.65%, from 95.00% to 97.31% and from 96.62%

to 97.45%, respectively. The average JDR of body

joints obtained by Mview-Joints improves the results

of Epipolar transformer (He et al., 2020) about 1%,

and with respect to Cross-View fusion (Qiu et al.,

2019) about 3% approximately.

Additionally, median Euclidean distance error is

used to evaluate the accuracy of prediction of the pro-

Human Pose Estimation through a Novel Multi-view Scheme

859

Figure 3: Challenging poses, the multi-view scheme takes advantage from the additional view with respect to the backbone—

single view—proposed by (Iskakov et al., 2019).

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

860

Figure 4: Comparison of Euclidean distance errors between the backbone approach (Iskakov et al., 2019) and the proposed

Mview-Joints for six different body joints.

posed multi-view scheme with respect to the CNN

backbone proposed by (Iskakov et al., 2019), whose

results are shown in Table 2. The body joints that

improvement signiﬁcant the accuracy are the elbow,

wrist, knee, nose, head. The median Euclidean dis-

tance errors for these joints improve by 15.88%,

4.32%, 8.46%, 18.11% and 50.53% respectively the

results obtained with CNN backbone proposed by

(Iskakov et al., 2019). Some challenging poses are

shown in Fig 3 where the multi-view scheme takes

advantage of the different views. The histograms of

accuracy of obtained body joints are shown in Fig. 4.

Most of the predicted 2D positions of body joints are

in the ranges [0-4] pixels and [4-8] pixels by using

Euclidean distance error for the proposed approach,

compared with the approach presented in (Iskakov

et al., 2019).

5 CONCLUSIONS

This paper addresses the challenging problem of the

human pose estimation when the joints are occluded.

A monocular network architecture that takes advan-

tage of multi-view scheme is proposed to accurately

estimate the human pose. This scheme is moti-

vated by the reduced information to predict more pre-

cisely occluded joints when only one view is used.

Experimental results and comparisons are provided

showing improvements on the obtained results. The

manuscript shows how estimated joints of other views

can help to estimate occluded joints more accurately.

The obtained precision of body joints is the base to

solve others related problems as 3D human pose es-

timation, action recognition among others. Future

work will be focused on extending the usage of multi-

view environments to leverage the geometry of the

scene, and thus, improve the 3D human pose.

ACKNOWLEDGEMENTS

This work has been partially supported by the

ESPOL projects EPASI (CIDIS-01-2018), TICs4CI

(FIEC-16-2018) and PhysicalDistancing (CIDIS-56-

2020); and the “CERCA Programme/Generalitat de

Catalunya”. The authors acknowledge the support of

CYTED Network: “Ibero-American Thematic Net-

work on ICT Applications for Smart Cities” (REF-

518RT0559) and the NVIDIA Corporation for the

donation of the Titan Xp GPU. The ﬁrst author

has been supported by Ecuador government under

a SENESCYT scholarship contract CZ05-000040-

2018.

REFERENCES

Cao, Z., Hidalgo, G., Simon, T., Wei, S.-E., and Sheikh, Y.

(2019). Openpose: realtime multi-person 2d pose esti-

mation using part afﬁnity ﬁelds. IEEE transactions on

pattern analysis and machine intelligence, 43(1):172–

186.

Carreira, J., Agrawal, P., Fragkiadaki, K., and Malik, J.

(2016). Human pose estimation with iterative error

Human Pose Estimation through a Novel Multi-view Scheme

861

feedback. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 4733–

4742.

Charco, J. L., Sappa, A. D., Vintimilla, B. X., and Vele-

saca, H. O. (2021). Camera pose estimation in multi-

view environments: From virtual scenarios to the real

world. Image and Vision Computing, 110:104182.

Charco, J. L., Vintimilla, B. X., and Sappa, A. D. (2018).

Deep learning based camera pose estimation in multi-

view environment. In 2018 14th International Confer-

ence on Signal-Image Technology & Internet-Based

Systems (SITIS), pages 224–228. IEEE.

Fang, H.-S., Xie, S., Tai, Y.-W., and Lu, C. (2017). Rmpe:

Regional multi-person pose estimation. In Proceed-

ings of the IEEE international conference on com-

puter vision, pages 2334–2343.

He, Y., Yan, R., Fragkiadaki, K., and Yu, S.-I. (2020).

Epipolar transformers. In Proceedings of the ieee/cvf

conference on computer vision and pattern recogni-

tion, pages 7779–7788.

Hofbauer, M., Kuhn, C. B., Meng, J., Petrovic, G., and

Steinbach, E. (2020). Multi-view region of inter-

est prediction for autonomous driving using semi-

supervised labeling. In 2020 IEEE 23rd Interna-

tional Conference on Intelligent Transportation Sys-

tems (ITSC), pages 1–6. IEEE.

Ionescu, C., Papava, D., Olaru, V., and Sminchisescu, C.

(2014). Human3.6m: Large scale datasets and pre-

dictive methods for 3d human sensing in natural envi-

ronments. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 36(7):1325–1339.

Iskakov, K., Burkov, E., Lempitsky, V., and Malkov, Y.

(2019). Learnable triangulation of human pose. In

Proceedings of the IEEE/CVF International Confer-

ence on Computer Vision, pages 7718–7727.

Newell, A., Yang, K., and Deng, J. (2016). Stacked hour-

glass networks for human pose estimation. In Euro-

pean conference on computer vision, pages 483–499.

Springer.

Qiu, H., Wang, C., Wang, J., Wang, N., and Zeng, W.

(2019). Cross view fusion for 3d human pose estima-

tion. In Proceedings of the IEEE/CVF International

Conference on Computer Vision, pages 4342–4351.

Remelli, E., Han, S., Honari, S., Fua, P., and Wang, R.

(2020). Lightweight multi-view 3d pose estimation

through camera-disentangled representation. In Pro-

ceedings of the IEEE/CVF Conference on Computer

Vision and Pattern Recognition, pages 6040–6049.

Sarmadi, H., Mu

noz-Salinas, R., Berb

ıs, M., and Medina-

Carnicer, R. (2019). Simultaneous multi-view camera

pose estimation and object tracking with squared pla-

nar markers. IEEE Access, 7:22927–22940.

Sun, K., Xiao, B., Liu, D., and Wang, J. (2019). Deep high-

resolution representation learning for human pose es-

timation. In Proceedings of the IEEE/CVF Conference

on Computer Vision and Pattern Recognition, pages

5693–5703.

Tang, C., Ling, Y., Yang, X., Jin, W., and Zheng, C. (2018).

Multi-view object detection based on deep learning.

Applied Sciences, 8(9):1423.

Tian, C., Xu, Y., Fei, L., Wang, J., Wen, J., and Luo, N.

(2019). Enhanced cnn for image denoising. CAAI

Transactions on Intelligence Technology, 4(1):17–23.

Tompson, J., Goroshin, R., Jain, A., LeCun, Y., and Bregler,

C. (2015). Efﬁcient object localization using convo-

lutional networks. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 648–656.

Toshev, A. and Szegedy, C. (2014). Deeppose: Human pose

estimation via deep neural networks. In Proceedings

of the IEEE conference on computer vision and pat-

tern recognition, pages 1653–1660.

Wei, S.-E., Ramakrishna, V., Kanade, T., and Sheikh, Y.

(2016). Convolutional pose machines. In Proceed-

ings of the IEEE conference on Computer Vision and

Pattern Recognition, pages 4724–4732.

Wu, M., Yue, H., Wang, J., Huang, Y., Liu, M., Jiang,

Y., Ke, C., and Zeng, C. (2020). Object detection

based on rgc mask r-cnn. IET Image Processing,

14(8):1502–1508.

Xiao, B., Wu, H., and Wei, Y. (2018). Simple baselines

for human pose estimation and tracking. In Proceed-

ings of the European conference on computer vision

(ECCV), pages 466–481.

Xie, H., Yao, H., Sun, X., Zhou, S., and Zhang, S. (2019).

Pix2vox: Context-aware 3d reconstruction from sin-

gle and multi-view images. In Proceedings of the

IEEE/CVF International Conference on Computer Vi-

sion, pages 2690–2698.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

862