Capsule network with using shifted windows for 3D human pose estimation

Liu, Xiufeng; Zhao, Zhongqiu; Tian, Weidong; Liu, Binbin; He, Hongmei

Abstract

3D human pose estimation (HPE) is a vital technology with diverse applications, enhancing precision in tracking, analyzing, and understanding human movements. However, 3D HPE from monocular videos presents significant challenges, primarily due to self-occlusion, which can partially hinder traditional neural networks’ ability to accurately predict these positions. To address this challenge, we propose a novel approach using a capsule network integrated with the shifted windows attention model (SwinCAP). It improves prediction accuracy by effectively capturing the spatial hierarchical relationships between different parts and objects. A Parallel Double Attention mechanism is applied in SwinCAP enhances both computational efficiency and modeling capacity, and a Multi-Attention Collaborative module is introduced to capture a diverse range of information, including both coarse and fine details. Extensive experiments demonstrate that our SwinCAP achieves better or comparable results to state-of-the-art models in the challenging task of viewpoint transfer on two commonly used datasets: Human3.6M and MPI-INF-3DHP.

• Journal Article Type: Article
• Acceptance Date: Feb 1, 2025
• Online Publication Date: Feb 10, 2025
• Publication Date: 2025-04
• Deposit Date: Mar 21, 2025
• Journal: Journal of Visual Communication and Image Representation
• Print ISSN: 1047-3203
• Publisher: Elsevier
• Peer Reviewed: Peer Reviewed
• Volume: 108
• Article Number: 104409
• DOI: https://doi.org/10.1016/j.jvcir.2025.104409