.png){kind=icon}
.png){kind=icon}
Return
From coarse to fine:Clip-cross hierarchical refinement network for 3D human pose estimation from monocular videos
.png){kind=icon}
PRE
AI
X
W
J
DOI：10.1016/j.patrec.2026.03.015.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In this research, we introduce the Clip-Cross Hierarchical Refinement Network (CCHRN), an innovative Transformer-based architecture designed for the estimation of 3D human pose and shape from monocular video inputs. This framework holds significant potential for applications in motion capture, augmented reality, and human-robot interaction. A primary challenge in this domain is the recovery of pixel-aligned 3D poses and shapes in the presence of inherent depth ambiguities. Current methodologies frequently encounter difficulties in capturing local motion discontinuities while simultaneously addressing global trends, often relying on decoupled kinematic constraints that undermine both temporal consistency and physical plausibility. To mitigate these challenges, CCHRN incorporates two essential components: the Clip-Cross Refinement Module (CCRM), which utilizes Transformer encoders to integrate multi-scale spatiotemporal features, and the Hierarchical Clip-Cross Regressor (HCCR), which enhances pose refinement by embedding kinematic priors derived from the SMPL model into the network. The integration of these modules allows CCHRN to effectively model joint-level dynamics and maintain temporal coherence. Evaluations conducted on the 3DPW and Human3.6 M datasets reveal competitive performance, with PA-MPJPE of 49.3 mm and 44.3 mm, respectively, thereby surpassing state-ofthe-art methods in achieving a balance between local precision and global consistency.
Keywords:
3D human pose estimation
Monocular video
Temporal modeling
Transformer
Kinematic constraints
Journal
IF:
3.3
Papers:
7.8K
Citations:
1.6W
Organization
Cited Papers