.png){kind=icon}
返回
Relation-balanced graph convolutional network for 3D human pose estimation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.imavis.2023.104841.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Graph convolutional networks (GCNs) have been applied to 2D-to-3D human pose estimation (HPE) and have shown encouraging performance. However, existing GCNs model the relations between joints via individual kernels, which can be overly flexible and fail to capture common relational patterns due to the symmetric nature of the human body. Although some GCNs share kernels to capture common relations, the unified way for all neighbors limits relational diversity to some extent. In order to balance the diversity and commonality of relations, we conduct a comprehensive study of existing kernel-sharing strategies and propose a Relation-balanced Graph Convolutional Network (RbGC-Net). RbGC-Net introduces the Part-Specific Kernel-Sharing strategy (PSKS) that assigns kernels based on the semantic meanings of neighbors to establish specific relational patterns for different types of neighborhoods. Furthermore, RbGC-Net incorporates a Local-Global Feature Fusion module (LGFF) that extracts the local relations among joints and balances them with the final global relations to improve the interactions between joints. Compared with state-of-the-art methods for 3D HPE, our RbGC-Net achieves the optimal balance between model size and estimation errors. Results on two benchmark Human3.6 M and MPIINF-3DHP datasets demonstrate the excellent performance and strong generalization ability of our pure GCNbased method.
Keyword:
3D human pose estimation
Graph convolutional network
Relation modeling
Feature fusion
期刊
IF:
4.2
论文数:
4.1K
被引数:
6.7K
机构
引用论文
分享收藏Graph convolutional network with structure pooling and joint-wise channel attention for action recognition.png)
PATTERN RECOGNITION
PATTERN RECOGNITIONIF7.6
PREAI
分享收藏分享收藏分享收藏分享收藏Preschool children's vision screening in New Zealand: a retrospective evaluation of referral accuracyBMJ Open
OAAI
BMJ OpenIF0
OAAI分享收藏
查看更多内容