MXPose: Multiplex interactive learning for multi-view 3D Human Pose Estimation

Wanruo Zhang

doi:10.1016/j.patcog.2026.113993

MXPose: Multiplex interactive learning for multi-view 3D Human Pose Estimation

2026-09-29

1

PRE

AI

DOI：10.1016/j.patcog.2026.113993

Original

Original request for help

Save

Abstract

En 中文

Multi-view 3D Human Pose Estimation (HPE) effectively alleviates occlusion and depth ambiguity. However, existing approaches often rely on precise camera calibration or monolithic modeling strategies, limiting robustness in complex scenarios. In this paper, MXPose is presented for multi-view 3D human pose estimation, which is a multiplex interactive learning framework. This framework models interactions across spatial structure, view-wise geometry, temporal dynamics, and optimization stages. Specifically, a Multiplex Interactive Spatial Encoder (MXS) is designed to realize spatial multiplex interaction by jointly modeling joint-level, part-level, and body-level representations through hierarchical dynamic hypergraphs. To facilitate view-wise interaction, a Cross-View Fusion (CVF) module is utilized to aggregate multi-view information into a unified geometric representation. Furthermore, a Multiplex Interactive Temporal Module (MXT) is introduced to integrate long-range temporal attention with adaptive multi-scale convolution and high-order kinematic cues. Finally, a Progressive Multi-Stage Supervision (PMSS) strategy is employed to enable interaction among geometric, structural, and projective objectives, guiding stable optimization from coarse initialization to refined alignment. Extensive experiments on Human3.6M, MPI-INF-3DHP, and Ski-Pose benchmarks demonstrate that MXPose consistently achieves state-of-the-art performance, validating its robustness in challenging, uncalibrated multi-view settings. Code is available at: https://github.com/Jokerealm/MXPose.

Keywords:

Multi-view 3D Human Pose Estimation

Multiplex interactive learning

Dynamic hypergraph

Temporal modeling

Progressive supervision

MXPose: Multiplex interactive learning for multi-view 3D Human Pose Estimation

Abstract

Journal

Organization

Cited Papers