Research on Human Pose Estimation Based on 2D and 3D

Classification

Xu Ji

, Can Liu

and Lingyue Zeng

Tianjin Institute of Software Engineering, Tianjin University of Technology, Tianjin, China

School of Computing and Data Science, Xiamen University Malaysia, Kuala Lumpur, Malaysia

Computer Science, University of Liverpool, Liverpool, U.K.

Keywords: Human Pose Estimation, Top-Down, Bottom-Up, Heatmap, Regression.

Abstract: Human pose estimation (HPE) refers to the use of computer vision technology and machine learning methods

to detect and analyse human pose information from images or videos. The early mainly focused on solving

single person pose estimation problems. It treats the problems as classification or regression problems using

a global feature. However, such methods are not accurate enough and only suitable for simple scenarios. The

application of HPE estimation based on deep learning is very promising and can provide many important

advantages and innovations in related field. Therefore, this paper mainly focuses on several methods and

technical advances for 2D and 3D HPE based on deep learning. This paper aims to introduce the technique

and compare the characteristics of different categories. Besides, it covers various datasets since 2015 and

different metrics for evaluating methods performance and illustrating considerable potential for future

development. Finally, this paper discusses the advantages and limitations through comparison of methods.

With the development of computer processing ability, the future human pose estimation tasks will make more

progress in improving accuracy, expanding application scenarios and optimizing efficiency.

1 INTRODUCTION

Human pose estimation (HPE) involves refers to

recognizing different parts of human body and

constructing representations such as body skeletons

from inputs like images and videos. Over the past

decade, this field has attracted growing interest and

has been applied across various domains (Zheng,

2023). Deep learning-based method has significantly

outperformed traditional methods in Human Pose

Estimation, primarily owing to three key factors: the

abundance of large-scale datasets, the enhanced

representation power of deep neural networks, and

the advent of high-performance hardware such as

Graphics Processing Unit (GPU) platforms (Lan,

2022). It stands out in its ability to process sports

analysis, computer-generated image, and so on,

thereby simplifying tasks like action recognition and

human tracking which emphasizes its potential as a

pivotal tool in computer vision.

https://orcid.org/0009-0008-6166-6726

https://orcid.org/0009-0002-6220-4079

https://orcid.org/0009-0004-0731-1001

2D HPE refers to the information of human pose

on two-dimensional image plane inferred from image

or video. The main tasks include detecting key points

of the human body (such as the head, shoulders,

elbows, wrists, knees, etc.) and estimating their

position and connection in the image. This approach

typically does not take depth information into account

and instead focuses on locating the location of key

points in the human body in a two-dimensional image

(Ben, 2021).

Beyond 2D human pose estimation research,

experimenters’ work extends to 3D HPE technology.

3D HPE refers to the estimation of human pose

information in 3D space from images or videos.

Different from two-dimensional pose estimation,

three-dimensional pose estimation not only focuses

on the position of key points in the image plane, but

also considers the three-dimensional coordinates of

these key points in space. Its main goal is to restore

the posture of the human body in three-dimensional

Ji, X., Liu, C. and Zeng, L.

Research on Human Pose Estimation Based on 2D and 3D Classiﬁcation.

DOI: 10.5220/0013234200004558

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management (MLSCM 2024), pages 99-105

ISBN: 978-989-758-738-2

space, including joint Angle, body orientation, spatial

position of the body and other information. Therefore,

3D pose estimation may be more challenging than 2D

(Ben, 2021).

To date, a wide range of research has been

conducted by scientists to explore the vast potential

of 2D HPE technology in computer vision. In 2021,

experimenters discuss the motion recognition

algorithm based on 2D method as well as using

convolutional neural network (CNN) for action

classification training (Chen, 2021). The results show

that the motion recognition algorithm based on 2D

HPE can improve the accuracy of motion recognition

to a certain extent and show its potential and

advantages in practical applications. Apart from 2D

HPE research, experimenters have expanded their

efforts into 3D HPE technology. In 2021, an

experimental team use a red-green-blue Depth Image

(RGBD) camera for 3D HPE and proved that this

method can help the system reconstruct the human

body's posture in 3D space more accurately (Fang,

2021). A serious of related approaches have been

tested and validated on publicly available data sets

and demonstrated superior pose estimation results for

3D HPE.

In this paper, the goal is to present a

comprehensive review for both 2D and 3D HPE

based on deep learning and research papers from

2015 to 2024 are covered. The 2D human pose

estimation calibration and key point visibility

classification technique, as well as a Wi-Fi-based

technique are introduced. In further discussion, this

paper deeply explores the 3D HPE technology based

on self-supervised learning and time encoder and

regression decoder. For the experimental section, the

data sets used to evaluate the various methods are

detailed. A variety of evaluation metrics are proposed

to evaluate the performance of different methods and

attain straightforward conclusion. Finally, through

the comparison of data analysis and experimental

results, this paper discusses the advantages and

limitations of the various methods and explore

possible challenges and solutions in future research.

2 METHOD

2.1 WiFi Signals to Estimate 2D

Human Poses

2.1.1 Method Introduction

In this experiment, 2D estimation of human body is

performed through Wireless Fidelity (Wi-Fi), but the

information contained in Wi-Fi signals is limited, so

it is necessary to extend the channel state information

(CSI) sequence and integrate spatial information and

dynamic information by using an air-frequency

encoder. To enable the model to focus on specific

channels, an evolutionary attention module is

designed. The model outperforms the current state-of-

the-art methods in PCK@20 by at least 16%.

2.1.2 Method Analysis

Wi-Fi signals to estimate 2D human poses can

capture human posture without wearing devices,

which can be widely used and promoted because it is

based on Wi-Fi. But precision is harder to guarantee

than sensors and cameras. In the processing of the

human posture room, a large amount of data needs to

be processed, so it will increase the cost and

complexity.

2.2 3D Human Pose Estimation Based

on Self-Supervision

2.2.1 Method Introduction

The method adopts Pose ResNet convolutional neural

architecture, which is based on ResNet architecture.

It is divided into 2D Pose CNN and 3D Pose CNN.

The two modules are composed of feature extraction

module and critical joint detection module. The

model can achieve self-supervised learning and

construct supervised information. The mechanism

adopted is Convolutional Block Attention Module

(CBAM), which can select the more important

information from all the information and suppress the

useless information. Using the CBAM module, the

information of the image can be obtained, and the

field of perception can be expanded, thus making the

detection more accurate.

2.2.2 Method Analysis

The advantage of this method is that 3D data is not

required, and 2D data can be used for training. Strong

generalization ability. Reliance on external

monitoring signals is reduced through self-

monitoring. This method can also be used in

combination with other technologies and has high

scalability. However, the performance of this method

in high-precision 3D attitude estimation is weaker

than that of traditional supervised learning methods.

If the data in the training data is biased, the self-

monitoring method will also be affected.

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

100

2.3 Three-Dimensional Human Pose

Estimation

2.3.1 Method introduction

The method uses Skinned Multi-Person Linear

(SMPL) to evaluate and model human posture. The

model is parameterized based on linear algebra. It

uses low-dimensional parameters to display

information about human posture and shape. The

framework of this method is divided into three parts,

which are linear encoding of data, time encoder and

motion regression decoder. Inertial Measurement

Unit (IMU) Feature Extractor enhances contract and

time capture by combining two consecutive frames.

Pose Triplet: Pose Triplet to transform the 2d pose

sequence into 3d output. The advantage is that it does

not rely on 3d data. The disadvantage is that 3d poses

cannot be directly estimated from 2d images.

Approach based on the encoder–decoder framework:

Individual frames are transformed into feature vectors

by Temporal Convolutional Encoder, Kinematic

Regression Decoder is used to predict the motion of

an object.

2.3.2 Method Analysis

Compared with the two-dimensional method, Three-

Dimensional Human Pose Estimation is more

abundant and accurate, which can make the generated

figures more real and has been widely used now.

However, the complexity of this method is high, the

demand for data is large, and the propagation error

will occur because of the reconstruction of three-

dimensional information.

3 EXPERIMENTS

This section describes the more widely used 2D and

3D datasets by 2024, as well as the common

evaluation metrics used for the different methods.

The performance results of these methods on different

datasets will also be presented.

3.1 Datasets

3.1.1 2D HPE Datasets

Deep learning-based methods are typically applied to

large amounts of training data, and this part will

introduce four mainly used 2D human pose datasets,

which are Max Planck Institute for Informatics

(MPII) (Andriluka,2014), Microsoft Common

Objects in Context (COCO) (Lin,2014), PoseTrack

(Andriluka,2018), Human-Art (Ju, 2023).

MPII dataset is a dataset to provide a high-quality

and diverse source of human pose annotations

including neck, head, elbow, which is important in

HPE.

COCO Dataset is the most widely used 2D dataset

including 200,000 labelled subjects with 80

categories such as people, animals and 330,000

images.

PoseTrack dataset is the datasets applied in the

video HPE task, with massive keypoints tracking.

Different scenarios include in the dataset such as

human stand overlapping resulting in occlusion of

body parts. Each human body is labeled with 15

major keypoints that describe the structure of human

body.

Human-Art is also widely used in 2D methods.

Comparison between these datasets is presented in

Table 1.

Table 1: 2D HPE Datasets. S: Single-person M: Multi-person.

Image-based datasets

Name Year S/M Joints

umber of ima

es Evaluation

protocol

Train Tes

MPII (Andriluka,2014) 2014 M

29k

3.8

12 k

1.7

PCKh

mAp

COCO2016 (Lin,2014) 2016 M 17 45

COCO2017 (Lin,2014) 2017 M 17 64

Human-Art (Ju, 2023) 2023 M 17 10

mAP

Video-based datasets

PoseTrack2017 (Andriluka,2018) 2017 M 15 250 214 mAP

PoseTrack2018 (Andriluka,2018) 2018 M 15 593 375 mAP

Research on Human Pose Estimation Based on 2D and 3D Classiﬁcation

101

Table 2: 3D HPE Datasets. S: Single-person M: Multi-person.

Datase

Yea

Capture s

stem Environmen

Includin

S M

AGORA

(Patel,2021)

2021 High-quality textured

scans

Indoor

14K images,

173K individual

Yes Yes

DensePose-COCO

(Güle

, 2018)

2018 Manual annotation

ali

nment with 3D models

Indoor

50K images, 5M

correspondences

Yes Yes

MuPoTS-3D

(Mehta,2018)

2018 Multi-View Marker-less Indoor and

outdoo

8 subjects, 8k

frames

Yes Yes

Human3.6M

(Ionescu,2014)

2014 Vicon Indoor 11 subjects

3.6M frames

Yes No

3.1.2 3D HPE Datasets

In 3D HPE datasets, Vicon and MoCap are required

to be able to accurately acquire 3D coordinates.

Human3.6 M (Ionescu, 2014) is the dataset covers

17 daily activities observed from 4 different

viewpoints. The dataset also utilizes a Vicon optical

motion capture system to acquire high-precision 3.6

million 3D pose data. It is the most important for 3D

HPE.

MuPoTS-3D (Multi-person Pose Tracking in 3D)

dataset is a 3D dataset specialized for multi-person

pose estimation and tracking. It contains 5 indoor

scenes and 15 outdoor scenes.

DensePose-COCO (Güler, 2018) and AGORA

(Patel, 2021) also included. Comparison between

these datasets is presented in Table 2.

3.2 Evaluation Metrics

3.2.1 2D HPE Evaluation Metrics

Object Keypoint Similarity (OKS) is a mertric

commonly used in COCO datasets to measures the

similarity between predicted keypoints and ground

truth keypoints.

The Average Precision (AP) metric is a measure

of the accuracy of critical point detection based on

precision and recall.

Percentage of Correct Keypoints (PCK) is used to

measures 2D HPE method by checking if predicted

keypoints fall within a threshold distance from

ground truth keypoints, commonly 50% of the head

segment length (PCKh@0.5).

3.2.2 3D HPE Evaluation Metrics

Area Under the Curve (AUC) plots the curve and

calculates the area under the curve by calculating the

keypoint correctness at multiple 3D PCK thresholds.

It provides a summary of the overall performance of

the model under different thresholds.

MPJPE (Mean Per Joint Position Error) is using

the Euclidean distance to estimate the accuracy of the

method as follows: 𝐽



is the truth keypoints, 𝐽



∗

is the

estimated keypoints.

𝑀𝐵𝐽𝑃𝐸 =

𝑁



‖

𝐽



−

𝐽



∗

‖









)

PA-MPJPE, is the MPJPE, where Procrustes

alignment is used to remove global rotation,

translation between the predicted and ground truth 3D

position to reduce the impact form the global factors.

3.3 Performance for Each Method

3.3.1 2D Method

Table 3 shows the results and performance of many

2D methods on the COCO dataset, analyzed by

comparing their AP values, backbone settings, and

input sizes. The different 2D HPE methods are

categorized into two types: top-down and bottom-up

to compare the performance of the two methods. In

general, the top-down method has better performance

than the bottom-up method in terms of accuracy of

pose estimation because the top-down method first

identifies all the people in the scene and then predicts

the location of the keypoints for each of them using

the single-person HPE method. Since top-down

methods in this category do not suffer from

performance degradation due to high background

complexity and high similarity between background

and people. Moreover, bottom-up methods perform

better and spend less time to finish the HPE task,

compare with top-down methods, because bottom-up

methods are advanced keypoint detection for the

characters in the scene and then use keypoints

association strategies such as Part Affinity Fields,

Pose Residual Network to categorize them into single

poses. The use of a transformer backbone in the top-

down approach improves the efficiency of the HPE

task without drastically decreasing its accuracy. In the

bottom-up method, the use of the HRNet-W32

backbone provides the best performance

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

102

Table 3: 2D Methods Performance in COCO Datasets with AP Measure. T: Top-down; B: Bottom-up.

Performance Data on COCO

Yea

Metho

Backbone Input size AP.5 AP(M) AP(L) Params(M) FLOPs(G)

2021

2022

(Yang,2021)

(Ma,2022)

HRNet-W48

Transforme

256x192

92.2

89.8

71.3

71.7

81.1

82.1

17.5

20.8

21.8

8.7

2020

2021

2022

(Jin,2020)

(Luo,2021)

(Wang,2022)

Hourglass

HRNet-W48

HRNet-W32

512x512

640x640

85.1

90.7

91.2

62.7

67.8

68.3

74.6

77.7

79.3

63.8

154.6

Table 4: 3D Methods Performance on the Human3.6M Dataset.

Skeleton-only methods MPJPE PA-MPJPE

Approaches Year Method Input Params(M) Average_1 Average_2

Direct

2017

2018

(Pavlakos,2017)

(Pavlakos,2018)

Image

Ima

71.9

56.3

52.0

41.7

Lifting

2020

2022

(Wang,2020)

(Zhan

,2022)

Video

42.5

40.9

32.7

32.5

Human mesh recover

methods MPJPE PA-MPJPE

Mesh Output

2021

2022

(Lin,2021)

(Cho,2022)

Image

Ima

230

51.2

52.1

34.5

33.6

2020

2021

(Kocabas,2020)

(Choi,2021)

Video

123

65.5

62.3

41.2

41.1

improvement, but is still lower than the top-down

method.

3.3.2 3D Method

Table 4 shows the results and performance of 3D

single-person HPE methods applied to the

Human3.6M dataset and categorized into two groups.

Almost all Single-person HPE methods are able to

accurately estimate the 3D human pose on

Human3.6M. However, they are affected by the

limitations of the Human3,6M dataset, such as the

lack of rich number and variety of character

movements, actor body types and environment

samples. When these methods are applied to estimate

3D human poses in more complex field scenarios,

performance degradation problems occur.

Meanwhile, estimating 3D human pose in video is

superior to estimating human pose estimation through

images, which exists with smaller MPJPE and higher

accuracy.

For Skeleton-only Methods, these methods

perform well with lower complexity, while the video

input 2D-to-3D lifting methods outperform the direct

estimation method overall by taking the support of a

high-performance 2D pose detector. For Human mesh

recovery methods, these methods suffer from high

Mean Per Joint Position Error (MPJPE) by using the

SMPL to recover, probably because of their high

mesh recovery complexity, which requires

simultaneous optimization and regression of multiple

objectives such as pose, shape, and surface details.

This complexity may lead to the gradual

accumulation of errors during the mesh

reconstruction process, which results in high MPJPE.

MPJPE is a limitation in evaluating such methods.

4 DISCUSSION

In 2D multi-person HPE, top-down methods have

higher accuracy but are slower, while bottom-up

methods are faster but relatively less accurate. To

balance speed and accuracy, a hierarchical detection

strategy can be adopted. First, the bottom-up method

is used to quickly detect the approximate positions of

all people, and then the top-down method is used to

finely locate the key points for each person.

Additionally, parallel computing techniques can be

used in both bottom-up and top-down methods,

accelerating processing speed through multi-GPU or

distributed computing. Finally, the key point

association strategy in bottom-up detection can be

improved, for example, by using graph neural

networks (GNNs) for key point association,

enhancing the accuracy of key point assembly and

thus improving overall accuracy.

The MPJPE of 3D skeleton and mesh recovery

methods is relatively high, primarily because these

methods regress both pose parameters and shape

parameters simultaneously, leading to error

Research on Human Pose Estimation Based on 2D and 3D Classiﬁcation

103

accumulation. To reduce error accumulation, a multi-

task learning framework can be adopted to separately

handle pose parameters and shape parameters. By

sharing some network layers, the model can

complement each other during the learning process,

improving overall accuracy. For example, a joint

network for pose estimation and shape reconstruction

can be designed, with each having its own

independent loss function. Additionally, the mesh

reconstruction algorithm can be improved to reduce

regression errors. Finally, a Generative Adversarial

Network (GAN) model can be used to improve the

quality of reconstructed meshes.

5 CONCLUSIONS

This paper introduces the current mainstream

methods from 2D and 3D respectively. The two-

dimensional human body is estimated based on Wi-Fi

signal, and the CSI sequence is integrated by spatial

encoder. Based on self-supervised 3D human Pose

estimation, the Pose ResNet convolutional neural

architecture is divided into 2D and 3D modules to

detect features and key joints. 3D human pose

estimation based on encoder and regression decoder

enhances the accuracy of human pose estimation by

adding both time and space.

At present, 2D motion capture can be divided into

camera-based visual capture, hand-drawn animation,

motion capture software, sensor capture. These

methods have their own characteristics and use cases.

In future research, it can be optimized for real-time

interaction. Optimize the real-time performance of

2D motion capture, improve the interactivity of the

system, and enable users to operate and edit more

conveniently.

At present, the mainstream 3D motion capture

method can be divided into optical camera systems,

inertial measurement unit, depth cameras, optical

tracking system, laser scanning system. These

methods have been widely used in different fields. In

future research, it can be combined with visual data,

inertial data, depth data, etc., so as to enhance the

accuracy of capture.

AUTHORS CONTRIBUTION

All the authors contributed equally and their names

were listed in alphabetical order.

REFERENCES

Ben Gamra, M., & Akhloufi, M. A., 2021. A review of deep

learning techniques for 2D and 3D human pose

estimation. Image and Vision Computing, 114, 104282.

https://doi.org/10.1016/j.imavis.2021.104282

Fang, Z., Wang, A., Bu, C., & Liu, C., 2021. 3D Human

Pose Estimation Using RGBD Camera. 2021 IEEE

International Conference on Computer Science,

Electronic Information Engineering and Intelligent

Control Technology (CEI).

https://doi.org/10.1109/cei52496.2021.9574486

Lan, G., Wu, Y., Hu, F., & Hao, Q., 2022. Vision-Based

Human Pose Estimation via Deep Learning: A Survey.

IEEE Transactions on Human-Machine Systems, 1–16.

https://doi.org/10.1109/thms.2022.3219242

Yu, C., Chen, W., Li, Y., & Chen, C., 2021. Action

Recognition Algorithm based on 2D Human Pose

Estimation Method. IEEE Xplore.

https://doi.org/10.23919/CCC52363.2021.9550204

Zheng, C., Wu, W., Chen, C., Yang, T., Zhu, S., Shen, J.,

Nasser Kehtarnavaz, & Shah, M., 2023. Deep

Learning-Based Human Pose Estimation: A Survey.

https://doi.org/10.1145/3603618

Andriluka, M., Pishchulin, L., Gehler, P., & Schiele, B.,

2014. 2D human pose estimation: New benchmark and

state of the art analysis. 2014 IEEE Conference on

Computer Vision and Pattern Recognition, 3686-3693.

https://doi.org/10.1109/CVPR.2014.471

Andriluka, M., Iqbal, U., Insafutdinov, E., Pishchulin, L.,

Milan, A., Gall, J., & Schiele, B., 2018. PoseTrack: A

Benchmark for Human Pose Estimation and Tracking.

IEEE Xplore.

https://doi.org/10.1109/CVPR.2018.00542

Ju, X., Zeng, A., Wang, J., Xu, Q., & Zhang, L., 2023.

Human-Art: A versatile human-centric dataset bridging

natural and artificial scenes. arXiv preprint.

https://doi.org/10.48550/arxiv.2303.02760

Cho, J., Youwang, K., & Oh, T.-H., 2022. Cross-Attention

of Disentangled Modalities for 3D Human Mesh

Recovery with Transformers. ArXiv.org.

https://arxiv.org/abs/2207.13820

Choi, H., Moon, G., Chang, J. Y., & Lee, K. M., 2020.

Beyond Static Features for Temporally Consistent 3D

Human Pose and Shape from a Video. ArXiv.org.

https://arxiv.org/abs/2011.08627

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

2014. Human3.6M: Large Scale Datasets and Predictive

Methods for 3D Human Sensing in Natural

Environments. IEEE Transactions on Pattern Analysis

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

https://doi.org/10.1109/TPAMI.2013.248

Jin, S., Liu, W., Xie, E., Wang, W., Qian, C., Ouyang, W.,

& Luo, P., 2020. Differentiable Hierarchical Graph

Grouping for Multi-Person Pose Estimation. ArXiv.org.

https://arxiv.org/abs/2007.11864

Kocabas, M., Athanasiou, N., & Black, M. J., 2020. VIBE:

Video Inference for Human Body Pose and Shape

Estimation. ArXiv.org. https://doi.org/10.48550/arXiv.

1912.05656

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

104

Lin, K., Wang, L., & Liu, Z., 2021. Mesh Graphormer. 2021

IEEE/CVF International Conference on Computer

Vision (ICCV). https://doi.org/10.1109/iccv48922.

2021.01270

Luo, Z., Wang, Z., Huang, Y., Tan, T., & Zhou, E., 2020.

Rethinking the Heatmap Regression for Bottom-up

Human Pose Estimation. ArXiv (Cornell University).

https://doi.org/10.48550/arxiv.2012.15175

Ma, H., Wang, Z., Chen, Y., Kong, D., Chen, L., Liu, X.,

Yan, X., Tang, H., & Xie, X., 2022. PPT: token-Pruned

Pose Transformer for monocular and multi-view

human pose estimation. ArXiv.org. https://arxiv.o

rg/abs/2209.08194

Mahmood, N., Ghorbani, N., Troje, N. F., Pons-Moll, G.,

& Black, M. J., 2019. AMASS: Archive of Motion

Capture as Surface Shapes. https://doi.org/10.1109/

iccv.2019.00554

Mehta, D., Oleksandr Sotnychenko, Mueller, F., Xu, W.,

Sridhar, S., Pons-Moll, G., & Theobalt, C., 2018.

Single-Shot Multi-person 3D Pose Estimation from

Monocular RGB. https://doi.org/10.1109/3dv.

2018.00024

Patel, P., Huang, C.-H. P., Tesch, J., Hoffmann, D. T.,

Tripathi, S., & Black, M. J., 2021. AGORA: Avatars in

Geography Optimized for Regression Analysis.

ArXiv.org. https://arxiv.org/abs/2104.14643

Pavlakos, G., Zhou, X., & Daniilidis, K., 2018. Ordinal

Depth Supervision for 3D Human Pose Estimation.

ArXiv.org. https://arxiv.org/abs/1805.04095

Pavlakos, G., Zhou, X., Derpanis, K. G., & Daniilidis, K.,

2017. Coarse-to-Fine Volumetric Prediction for Single-

Image 3D Human Pose. 2017 IEEE Conference on

Computer Vision and Pattern Recognition (CVPR).

https://doi.org/10.1109/cvpr.2017.139

Tang, W., Yu, P., & Wu, Y., 2018. Deeply Learned

Compositional Models for Human Pose Estimation.

Lecture Notes in Computer Science, 197–214.

https://doi.org/10.1007/978-3-030-01219-9_12

Wang, H., Zhou, L., Chen, Y., Tang, M., & Wang, J., 2022.

Regularizing Vector Embedding in Bottom-Up Human

Pose Estimation. Lecture Notes in Computer Science,

107–122.https://doi.org/10.1007/978-3-031-20068-7_7

Wang, J., Yan, S., Xiong, Y., & Lin, D., 2020. Motion

Guided 3D Pose Estimation from Videos. ArXiv.org.

https://arxiv.org/abs/2004.13985

Yang, S., Quan, Z., Nie, M., & Yang, W., 2021. TransPose:

Keypoint Localization via Transformer. ArXiv.org.

https://doi.org/10.48550/arXiv.2012.14214

Zhang, J., Tu, Z., Yang, J., Chen, Y., & Yuan, J., 2022.

MixSTE: Seq2seq Mixed Spatio-Temporal Encoder for

3D Human Pose Estimation in Video. 2022 IEEE/CVF

Conference on Computer Vision and Pattern

Recognition (CVPR). https://doi.org/10.1109/cvpr5268

8.2022.01288

Güler, R. A., Neverova, N., & Kokkinos, I., 2018.

DensePose: Dense Human Pose Estimation In The

Wild. ArXiv:1802.00434 [Cs]. https://arxiv.org/abs/

1802.00 434

Kocabas, M., Athanasiou, N., & Black, M. J., 2020. VIBE:

Video inference for human body pose and shape

estimation. In Proceedings of the IEEE Conference on

Computer Vision and Pattern Recognition (CVPR).

Choi, H., Moon, G., Chang, J. Y., & Lee, K. M., 2021.

Beyond static features for temporally consistent 3D

human pose and shape from a video. In Proceedings of

the IEEE Conference on Computer Vision and Pattern

Recognition (CVPR).

Bao, W., Ma, Z., Liang, D., Yang, X., & Niu, T., 2023. Pose

ResNet: 3D Human Pose Estimation Based on Self-

Supervision. Sensors (Basel, Switzerland), 23(6), 3057.

https://doi.org/10.3390/s23063057

Liao, X., Dong, J., Song, K., & Xiao, J., 2023. Three-

Dimensional Human Pose Estimation from Sparse

IMUs through Temporal Encoder and Regression

Decoder. Sensors (Basel, Switzerland), 23(7), 3547.

https://doi.org/10.3390/s23073547

Research on Human Pose Estimation Based on 2D and 3D Classiﬁcation

105