• Classified by Topic • Classified by Publication Type • Sorted by Date • Sorted by First Author Last Name • Classified by Funding Source •
Baby Mesh Recovery Enables Infant Behavior Identification.
Charles Dardaman, Zifan
Xu, Peter Stone, Karen Adolph, Danyang Han, and Georgios Pavlakos.
In
The IEEE International Conference on Development and Learning (ICDL), September 2026.
Natural infant motor behavior, such as sitting, crawling, and walking, is highly variable and idiosyncratic, and often deviates from standard, adult-like postures and locomotion. This variability makes it challenging for automated tools to support precise 3D video analysis of infant behavior. Here, we present a system for accurate 3D motion capture of infants' moment-to-moment body posture and step-to-step locomotion using synchronized multi-view RGB video. Our method builds on recent advances in parametric human modeling by adapting them to the unique challenges of infant behavior. We use a parametric baby body model--a "mesh baby"-- to represent age-varying body shapes and develop a multi-view optimization approach that jointly recovers 3D body pose, shape, and motion. We demonstrate that our system provides accurate full-body 3D baby reconstruction and enables fine-grained analysis, including identification of posture, step counts, and gait timing. Our approach provides a scalable, non-intrusive, video-based tool for understanding infants' natural motor behavior and opens new avenues for early screening of developmental disorders and evaluation of clinical interventions. We release the code and results at https://geopavlakos.github.io/babymesh/.
@InProceedings{charles_dardaman_ICDL2026,
author = {Charles Dardaman and Zifan Xu and Peter Stone and Karen Adolph and Danyang Han and Georgios Pavlakos},
title = {Baby Mesh Recovery Enables Infant Behavior Identification},
booktitle = {The IEEE International Conference on Development and Learning (ICDL)},
year = {2026},
month = {September},
location = {Kyoto, Japan},
abstract = {Natural infant motor behavior, such as sitting, crawling, and walking, is highly variable and idiosyncratic, and often deviates from standard, adult-like postures and locomotion. This variability makes it challenging for automated tools to support precise 3D video analysis of infant behavior. Here, we present a system for accurate 3D motion capture of infants' moment-to-moment body posture and step-to-step locomotion using synchronized multi-view RGB video. Our method builds on recent advances in parametric human modeling by adapting them to the unique challenges of infant behavior. We use a parametric baby body model--a "mesh baby"-- to represent age-varying body shapes and develop a multi-view optimization approach that jointly recovers 3D body pose, shape, and motion. We demonstrate that our system provides accurate full-body 3D baby reconstruction and enables fine-grained analysis, including identification of posture, step counts, and gait timing. Our approach provides a scalable, non-intrusive, video-based tool for understanding infants' natural motor behavior and opens new avenues for early screening of developmental disorders and evaluation of clinical interventions. We release the code and results at https://geopavlakos.github.io/babymesh/.},
}
Generated by bib2html.pl (written by Patrick Riley ) on Tue Jul 14, 2026 00:00:18