JCN: Joint Constraint-Based Human Pose Refinement Networks
2D human pose estimation has essential applications in traffic prediction and human-computer interaction. We propose a pose refinement network for refining human pose features to improve human pose detection accuracy. We propose our approach to two critical problems in pose estimation, occlusion and...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10543053/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1849310511868411904 |
|---|---|
| author | Yuru Zhang Jiayuan Zhao Xiaodong Su Hongyan Xu Meijian Jin |
| author_facet | Yuru Zhang Jiayuan Zhao Xiaodong Su Hongyan Xu Meijian Jin |
| author_sort | Yuru Zhang |
| collection | DOAJ |
| description | 2D human pose estimation has essential applications in traffic prediction and human-computer interaction. We propose a pose refinement network for refining human pose features to improve human pose detection accuracy. We propose our approach to two critical problems in pose estimation, occlusion and background interference, for which the model often needs local information for inference. In contrast, for the background interference problem, the model needs more non-local semantic information to guide the model to make semantic judgments. Specifically, we propose Joint Constraint-Based Human Pose Refinement Networks, which explicitly and implicitly model the relationship between critical points through graph neural networks and self-attention mechanisms to capture local and non-local information of the pose features and refine the keypoints responses. In order to avoid mutual conflicts between explicit and implicit modeling, this paper chooses a parallel structure. It selects an appropriate feature fusion strategy to fuse parallel branch features to improve model robustness. Meanwhile, to make the model pay more attention to the hard-to-identify vital points, this paper adopts the focal loss function to optimize the model and improve the data’s long-tailed distribution. We validate our work on the publicly available dataset MPII with the MSCOCO dataset. The results show that our proposed method has the effect of refining the keypoints response, with a PCKh@0.5 of 92.51% on the official MPII validation dataset, and achieved a result of 78% AP on the official MSCOCO validation dataset, which is comparable to the performance of existing state-of-the-art models. |
| format | Article |
| id | doaj-art-56da3d0dd9a54ef6a4fd7d4cd25962e6 |
| institution | Kabale University |
| issn | 2169-3536 |
| language | English |
| publishDate | 2025-01-01 |
| publisher | IEEE |
| record_format | Article |
| series | IEEE Access |
| spelling | doaj-art-56da3d0dd9a54ef6a4fd7d4cd25962e62025-08-20T03:53:42ZengIEEEIEEE Access2169-35362025-01-0113688416885610.1109/ACCESS.2024.340785210543053JCN: Joint Constraint-Based Human Pose Refinement NetworksYuru Zhang0Jiayuan Zhao1https://orcid.org/0000-0002-3319-9509Xiaodong Su2https://orcid.org/0009-0005-9451-8410Hongyan Xu3https://orcid.org/0009-0000-9840-7383Meijian Jin4School of Computer and Information Engineering, Harbin University of Commerce, Harbin, ChinaSchool of Computer and Information Engineering, Harbin University of Commerce, Harbin, ChinaSchool of Computer and Information Engineering, Harbin University of Commerce, Harbin, ChinaSchool of Computer and Information Engineering, Harbin University of Commerce, Harbin, ChinaSchool of Computer and Information Engineering, Harbin University of Commerce, Harbin, China2D human pose estimation has essential applications in traffic prediction and human-computer interaction. We propose a pose refinement network for refining human pose features to improve human pose detection accuracy. We propose our approach to two critical problems in pose estimation, occlusion and background interference, for which the model often needs local information for inference. In contrast, for the background interference problem, the model needs more non-local semantic information to guide the model to make semantic judgments. Specifically, we propose Joint Constraint-Based Human Pose Refinement Networks, which explicitly and implicitly model the relationship between critical points through graph neural networks and self-attention mechanisms to capture local and non-local information of the pose features and refine the keypoints responses. In order to avoid mutual conflicts between explicit and implicit modeling, this paper chooses a parallel structure. It selects an appropriate feature fusion strategy to fuse parallel branch features to improve model robustness. Meanwhile, to make the model pay more attention to the hard-to-identify vital points, this paper adopts the focal loss function to optimize the model and improve the data’s long-tailed distribution. We validate our work on the publicly available dataset MPII with the MSCOCO dataset. The results show that our proposed method has the effect of refining the keypoints response, with a PCKh@0.5 of 92.51% on the official MPII validation dataset, and achieved a result of 78% AP on the official MSCOCO validation dataset, which is comparable to the performance of existing state-of-the-art models.https://ieeexplore.ieee.org/document/10543053/Attention mechanismfeature fusion strategygraph neural networkhuman pose estimationjoint constraintparallel structure |
| spellingShingle | Yuru Zhang Jiayuan Zhao Xiaodong Su Hongyan Xu Meijian Jin JCN: Joint Constraint-Based Human Pose Refinement Networks IEEE Access Attention mechanism feature fusion strategy graph neural network human pose estimation joint constraint parallel structure |
| title | JCN: Joint Constraint-Based Human Pose Refinement Networks |
| title_full | JCN: Joint Constraint-Based Human Pose Refinement Networks |
| title_fullStr | JCN: Joint Constraint-Based Human Pose Refinement Networks |
| title_full_unstemmed | JCN: Joint Constraint-Based Human Pose Refinement Networks |
| title_short | JCN: Joint Constraint-Based Human Pose Refinement Networks |
| title_sort | jcn joint constraint based human pose refinement networks |
| topic | Attention mechanism feature fusion strategy graph neural network human pose estimation joint constraint parallel structure |
| url | https://ieeexplore.ieee.org/document/10543053/ |
| work_keys_str_mv | AT yuruzhang jcnjointconstraintbasedhumanposerefinementnetworks AT jiayuanzhao jcnjointconstraintbasedhumanposerefinementnetworks AT xiaodongsu jcnjointconstraintbasedhumanposerefinementnetworks AT hongyanxu jcnjointconstraintbasedhumanposerefinementnetworks AT meijianjin jcnjointconstraintbasedhumanposerefinementnetworks |