TY - GEN
T1 - 3D POSE ESTIMATION BY GROUPED FEATURE FUSION AND MOTION AMPLITUDE ENCODING
AU - Peng, Jihua
AU - Zhou, Yanghong
AU - Mok, P. Y.
N1 - Funding Information:
This research is funded by the Laboratory for Artificial Intelligence in Design (Project Code: RP1-1) under the InnoHK Research Clusters, Hong Kong Special Administrative Region Government.
Publisher Copyright:
© MCCSIS 2022.All rights reserved.
PY - 2022/7
Y1 - 2022/7
N2 - 3D human pose estimation is challenging and converting it to a local pose estimation problem by dividing human body into different groups based on anatomical relationships can improve the accuracy of the resulting 3D pose estimation. Joint features of different groups are fused to predict complete pose information of entire body, in which a joint feature fusion scheme must be used for this purpose. However, the joint feature fusion adopted in existing methods has to learn a large number of parameters and is computational expensive. In this paper, we propose an optimized feature fusion (OFF) module, which requires fewer parameters and less calculations while ensures prediction accuracy. Moreover, we also propose a motion amplitude encoding (MAE) method to improve the prediction accuracy for small ranged movements. Experiments have shown that our method outperforms previous state-of-the-art results on Human3.6M dataset.
AB - 3D human pose estimation is challenging and converting it to a local pose estimation problem by dividing human body into different groups based on anatomical relationships can improve the accuracy of the resulting 3D pose estimation. Joint features of different groups are fused to predict complete pose information of entire body, in which a joint feature fusion scheme must be used for this purpose. However, the joint feature fusion adopted in existing methods has to learn a large number of parameters and is computational expensive. In this paper, we propose an optimized feature fusion (OFF) module, which requires fewer parameters and less calculations while ensures prediction accuracy. Moreover, we also propose a motion amplitude encoding (MAE) method to improve the prediction accuracy for small ranged movements. Experiments have shown that our method outperforms previous state-of-the-art results on Human3.6M dataset.
KW - Feature Fusion
KW - Human Pose Estimation
KW - Motion Amplitude
UR - http://www.scopus.com/inward/record.url?scp=85142376282&partnerID=8YFLogxK
M3 - Conference article published in proceeding or book
AN - SCOPUS:85142376282
T3 - 16th International Conference on Computer Graphics, Visualization, Computer Vision and Image Processing, CGVCVIP 2022, 8th International Conference on Connected Smart Cities, CSC 2022, 7th International Conference on Big Data Analytics, Data Mining and Computational Intelligence, BigDaCI 2022, and 11th International Conference on Theory and Practice in Modern Computing, TPMC 2022 - Held at the 16th Multi Conference on Computer Science and Information Systems, MCCSIS 2022
SP - 27
EP - 34
BT - 16th International Conference on Computer Graphics, Visualization, Computer Vision and Image Processing, CGVCVIP 2022, 8th International Conference on Connected Smart Cities, CSC 2022, 7th International Conference on Big Data Analytics, Data Mining and Computational Intelligence, BigDaCI 2022, and 11th International Conference on Theory and Practice in Modern Computing, TPMC 2022 - Held at the 16th Multi Conference on Computer Science and Information Systems, MCCSIS 2022
PB - IADIS Press
T2 - 16th International Conference on Computer Graphics, Visualization, Computer Vision and Image Processing, CGVCVIP 2022, 8th International Conference on Connected Smart Cities, CSC 2022, 7th International Conference on Big Data Analytics, Data Mining and Computational Intelligence, BigDaCI 2022, and 11th International Conference on Theory and Practice in Modern Computing, TPMC 2022 - Held at the 16th Multi Conference on Computer Science and Information Systems, MCCSIS 2022
Y2 - 19 July 2022 through 22 July 2022
ER -