This is the authors' abstract. We don't add key points for this paper.
Existing multi-view 3D human pose estimation methods heavily rely on precise extrinsic calibration, which significantly restricts their practical deployment in uncontrolled environments. To address this limitation, we propose an Extrinsic Parameter-free Multi-view 3D Human Skeleton Estimation (EFMK) framework with three technical contributions. First, a Local-Global Pose Embedding scheme is proposed to simultaneously capture the fine-grained joint dependencies while establishing cross-view correspondences. Second, a Spatial-View Joint Transformer architecture is developed with three dedicated components: (1) Feature Transformation Modulation generates adaptive modulation vectors for distinct tokens to model heterogeneous relationship patterns; (2) Prior Knowledge Enhancement systematically integrates human kinematic constraints and multi-view geometric priors into attention computation through structural topology encoding; (3) Spatial-View Joint Attention implements decoupled spatial-view attention computation followed by joint distribution modeling to capture hierarchical spatial-view dependencies. Third, a Bone-wise Reprojection-based Multi-view Aggregation mechanism is introduced to consolidate multiple 3D outputs into a single, higher-quality 3D pose for practical applications. Extensive experiments on three benchmarks demonstrate that our method achieves state-of-the-art performance while maintaining a compact model size. Code and results are available at https://github.com/Z-Z-J/EFMK.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: