Mitigating the Impact of Labeling Inaccuracies on 3D Human Body Reconstruction from Monocular Videos

Abstract This paper addresses the challenge of labeling inaccuracies in 3D human pose and shape reconstruction from monocular videos. Existing methods often rely on noisy pseudo ground truth, which introduces performance degradation such as jittering and drifting. To overcome these limitations, we p...

Full description

Saved in:
Bibliographic Details
Main Authors: Yupeng Hou, Guangping Zeng
Format: Article
Language:English
Published: Springer 2025-07-01
Series:International Journal of Computational Intelligence Systems
Subjects:
Online Access:https://doi.org/10.1007/s44196-025-00921-5
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract This paper addresses the challenge of labeling inaccuracies in 3D human pose and shape reconstruction from monocular videos. Existing methods often rely on noisy pseudo ground truth, which introduces performance degradation such as jittering and drifting. To overcome these limitations, we propose a confidence-aware framework grounded in biomechanics. Our method adopts the SKEL model to provide anatomically constrained pose representations, reducing dependency on imprecise annotations. We further introduce a conditional normalizing flow to model per-parameter uncertainty conditioned on visual and motion features. Additionally, a novel evaluation metric, confidence-weighted procrustes aligned MPJPE, is proposed to incorporate confidence scores into performance assessment. Extensive experiments show that our approach outperforms existing methods on multiple datasets in both accuracy and motion smoothness. It demonstrates strong robustness against noisy annotations, and confidence estimates align closely with actual prediction errors. Ablation studies validate the contributions of both biomechanical modeling and confidence learning. Overall, our framework provides a unified and robust solution for 3D human reconstruction in real-world settings.
ISSN:1875-6883