Page 63 - 2024F

56 UEC Int’l Mini-Conference No.53

Optimized Graph Neural Network

Approach for 3D Human Pose Estimation

Acevedo-Bringas Luis*, Hiroki Takahashi
The University of Electro-Communications
Department of Informatics
*a2440012@gl.cc.uec.ac.jp

Introduction

3D Human Pose Estimation (HPE) in videos aims to predict the pose joint locations of the human body
in a 3D space.

• Graph-based models: human body as graph, joints as nodes, spatial, temporal & relational learning.
• Transformer-based models: self-attention, long-range dependencies, global context.

We choose the graph-based models because they effectively model the human body's kinematic and
structural relationships, leading to more interpretable and computationally efficient pose estimation.

Objective: Develop a Graph-based model capable of refine and detect human key points with a trade-
off between accuracy and computational cost.

Methodology

Figure 1. Baseline architecture (GLA-GCN[1])

Metrics: Expected contributions
• MPJPE↓: Average Euclidian Distance
between the joints predicted and the Stage 1: Implement global and local
GT. attention for the key points.
• PCK↑: Percentage of Correct
predicted key points Stage 2: Implement human tracking.

Datasets: Stage 3: Search for the best trade-off
st
For the 1 , 2 nd and 3 rd stage the configuration between computation
experiments are conducted using the cost and accuracy. Figure 2.
Human3.6M and MPI-INF-3DHP datasets Global
rd
For the 4 stage the experiments will be Stage 4: Implement the algorithm for interaction of
using the AIT-Soccer-FIFA-Skeletal soccer games analysis. a key point[2].
dataset

References
[1] B. X. B. Yu, Z. Zhang, et. al, "GLA-GCN: Global-local adaptive graph convolutional network for 3D human pose estimation
from monocular video," in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2024
[2] Wang Ti, Hong Liu, et. al, Interweaved Graph and Attention Network for 3D Human Pose Estimation, in In IEEE
International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2023