Prior-guided decoupled convolutional fusion network for multi-view 3D human pose estimation

Abstract Multi-view 3D Human Pose Estimation (3D-HPE) achieves superior accuracy over single-view approaches, yet its reliance on precise camera calibration fundamentally restricts deployment in uncalibrated or dynamic environments. Moreover, existing methods typically entangle spatial, view-wise, and temporal modeling within a unified architecture, leading to computational redundancy and limited scalability. To address these challenges, we propose a Decoupled Spatio-Temporal Paradigm that fundamentally rethinks the architectural design of multi-view 3D-HPE. Rather than incrementally replacing individual modules, our paradigm decomposes the entire spatio-temporal modeling pipeline into three sequential and independent stages—Prior Enhancement, View Alignment, and Temporal Decoding—each governed by a dedicated architectural principle. Three key innovations distinguish our work. (i) At the prior enhancement stage, a kinematic-aware module explicitly injects anatomical constraints to refine 2D representations before any cross-view reasoning, thereby reducing the burden on subsequent fusion. (ii) At the view alignment stage, a calibration-free dual-branch module jointly learns semantic attention weights and geometric transformation matrices, enabling adaptive fusion of an arbitrary number of views without camera parameters. (iii) At the temporal decoding stage, we demonstrate that an architecture built entirely upon Large-Kernel Depth-wise Convolutions can serve as a complete replacement for Transformer-based temporal models, providing stronger inductive bias for motion continuity at a fraction of the computational cost. Without any camera calibration, our method achieves 25.0 mm MPJPE on Human3.6M, outperforming all existing calibration-free methods by 0.6–4.5 mm, with only 7.9 M parameters and 9.1 ms single-sample inference latency on a commodity GPU, demonstrating that decoupled spatio-temporal design can simultaneously advance accuracy, efficiency, and deployment flexibility.